Skip to main content
PLOS One logoLink to PLOS One
. 2026 Oct 1;21(10):e0358970. doi: 10.1371/journal.pone.0358970

Informed forecasting: Leveraging auxiliary knowledge to boost large language model performance on time series forecasting

Mohammadmahdi Ghasemloo 1, Alireza Moradi 2,*
Editor: Dinesh Kumar Nishad3
PMCID: PMC13630261  PMID: 42821624

Abstract

The rapid adoption of large language models (LLMs) has sparked growing interest in extending their capabilities beyond traditional natural language tasks, including applications in time series forecasting. This work explores the enhancement of time series forecasting with LLMs by incorporating time-dependent covariates. We propose a set of representative prompting strategies that span a wide range of formats while incorporating covariates, and then implement the proposed framework to evaluate performance across three real-world time series datasets in healthcare, service operations, and transportation. In our experiments, we use GPT-4o-mini and show that incorporating relevant covariate information through an appropriate prompt design can substantially improve forecasting accuracy, reducing prediction errors by up to 70% compared with the no-covariate baseline. Furthermore, the analysis reveals that both the choice of covariate and the prompt structure are critical, as poorly aligned configurations may degrade performance. Finally, we conduct sensitivity analyses to assess the effect of covariate integration in censored settings and quantify uncertainty, confirming that our method achieves statistically significant improvements over existing approaches. These findings underscore the potential of covariate integration in prompt design to bridge the gap between general-purpose LLMs and forecasting tasks.

1. Introduction

The emergence of Large Language Model (LLM) services has led to a rapid increase in their global usage. Platforms such as ChatGPT, for example, report more than 300 million weekly users, which underscores the widespread integration of these tools into daily life and professional workflows [1]. This rapid adoption has, in turn, catalyzed efforts to repurpose LLMs beyond core NLP settings.

LLMs originally designed for understanding and generating natural language have quickly expanded to a wide range of complex domains. Recent research has demonstrated their promising capabilities in areas such as robotics control, optimization, simulation-based reasoning, task-oriented dialog systems, and time series forecasting [2–8]. Building on this trend, we focus specifically on the role of LLMs in time series forecasting.

Time series forecasting plays a vital role in decision-making across a range of application domains. In healthcare, accurate forecasts of disease incidence can inform resource allocation, staffing, and public health interventions. In service operations, demand forecasting enables better capacity planning and scheduling. In finance, reliable forecasts support investment strategies and risk management. In power systems and industrial operations, data-driven models have similarly been applied to fault diagnosis and stability assessment, where timely detection of anomalies is critical for safe and efficient operation [9,10]. Given these stakes, it is natural to ask when and how general-purpose LLMs can contribute to time series forecasting in a useful and low-friction manner.

Despite the existence of specialized time series forecasting models, there are compelling reasons to explore the utilization of commonly used LLMs in this domain. First, LLMs offer a highly accessible and low-barrier alternative for non-expert users, enabling a broader community to perform forecasting tasks without modeling expertise. Second, LLM-based forecasting is particularly well-suited for use as an initial decision-support tool in operational settings such as healthcare (e.g., predicting patient inflow), call centers (e.g., estimating daily arrival volumes), and transportation (e.g., forecasting airline passenger demand), where fast, interpretable, and adaptive forecasts can assist in planning and resource allocation. Third, due to their fast inference and natural language interface, LLMs can be used to generate preliminary forecasts that serve as input models for high-fidelity decision tools, including simulation-based systems. These practical advantages motivate a closer look at design choices for LLM-based time series forecasting.

In classical time series forecasting, covariates such as calendar attributes, weather conditions, or economic indicators play a crucial role in improving predictive accuracy. However, unlike traditional forecasting methods, most LLM-based approaches rely on raw inputs and overlook other information available in the dataset [11]. Recent work has started to explore this direction. [11] augmented LLM inputs with simple temporal covariates such as data entry date. [5] emphasized the effect of external knowledge by embedding contextual attributes derived from the dataset itself. Similarly, [12] proposed a retrieval-augmented framework in which relevant financial indicators are selected and included in the input prompt.

Together, these efforts establish the value of utilizing side information but leave three questions open. First, there is limited guidance on how users should formulate effective prompts to incorporate covariates into the model. More specifically, what constitutes a good prompt for presenting covariate information in a way that enhances predictive performance without overwhelming or misleading the model? Second, a systematic approach is needed to evaluate the effect of covariate integration on the accuracy of forecasting. In other words, how can we determine which variables contribute meaningfully to predictive accuracy and how can this selection process be made robust across different datasets and forecast horizons? Third, considerations such as data leakage and computational cost are critical alongside forecast accuracy. Therefore, how can users effectively mitigate the risk of data leakage while also accounting for cost?

To address this gap, this paper introduces a pipeline for informed LLM-based time series forecasting that examines the role of time-dependent covariates available within the dataset. To ensure practical generality, we focus on covariates that are not tied to a specific dataset, chiefly calendar-derived attributes that can be defined for virtually any series, and we employ a validation-based framework to evaluate each prompt and covariate in terms of forecasting performance.

The contributions of this work are threefold. First, we propose a covariate-informed prompting framework that explicitly integrates time-dependent covariates into LLM-based forecasting prompts. Second, we introduce and compare a curated set of representative prompt formats to study how different ways of organizing covariate information affect forecast accuracy. Third, we develop a validation-based procedure for selecting prompt–covariate pairs, allowing users to determine whether a given covariate improves forecasting performance for a specific dataset and forecast horizon.

We provide a curated set of representative prompt formats tested on three real-world datasets. These experiments demonstrate how incorporating time-dependent covariates into prompt design can significantly improve forecast performance. We do not claim to have discovered the optimal prompt for all scenarios, nor that our approach outperforms traditional time series forecasting methods in general. Rather, we argue that when commonly used LLMs are used for forecasting, their performance can be significantly enhanced by incorporating covariates into the prompt design. Our objective, therefore, is to surface effective, reusable patterns for covariate integration rather than to replace specialized time series forecasting models. Fig 1 illustrates the end-to-end forecasting workflow.

Fig 1. Overview of the process of integrating covariates into the task of forecasting with LLMs.

Fig 1

A time-dependent covariate is integrated with raw time series data to generate forecasts through natural language interaction.

The remainder of this paper is organized as follows. Section 2 reviews related work on classical and modern time series forecasting methods as well as recent developments in LLMs for forecasting tasks. Section 3 introduces the proposed framework for covariate-informed LLM-based time series forecasting. Section 4 evaluates the effect of incorporating covariates on three real-world datasets, and finally, Section 5 summarizes the key findings and outlines directions for future research.

2. Related work

Time series forecasting has long been one of the most attractive and widely applicable areas of research for data-driven fields. Consequently, significant efforts have been made by the research community to advance this domain [13]. The potential applications of time series forecasting are vast and diverse, encompassing domains such as finance [14,15], healthcare [16], power systems [17], and supply chain management [18], among others. This wide range of use cases underscores the importance of time series forecasting in real-world operations and decision-making. As a result, numerous methodologies have been developed for time series prediction, including classical statistical approaches such as ARIMA [19–21]. For example, [19] applied ARIMA to forecast the Nigerian stock market, while [20] used it to forecast electricity demand with seasonal adjustments. Linear prediction models have also been extensively employed in forecasting tasks [22,23]. [22] presented a comprehensive review of linear and hybrid models in the prediction of energy systems, and [23] applied multiple linear regression to predict electricity demand in Italy.

With the advancement of computational capabilities, machine learning techniques have gained traction in time series forecasting [24,25]. [24] demonstrated the effectiveness of Random Forests in predicting electricity load with high-dimensional input, while [25] applied machine learning models to predict forest disturbances using remote sensing data. More recently, deep learning frameworks have emerged as powerful tools for modeling complex temporal patterns in time series data [14,26]. [14] reviewed deep learning approaches such as CNNs and LSTMs in financial time series forecasting, while [26] performed a comparative analysis of deep learning models in various forecasting tasks and highlighted their superior performance over traditional methods. In addition, [8] proposed an explainable deep learning architecture for multivariate time series forecasting that provides interpretability while maintaining high predictive accuracy.

On the other hand, LLMs have found significant applications in scientific research, expanding their role beyond their initial design [27]. Although the primary function of LLMs is to process, understand, and generate human-like text by identifying patterns and relationships in vast datasets [28], recent research has explored their potential for applications that go beyond their original purpose. Several recent works have initiated this exploration. For instance, [2] introduced Time-LLM, a framework that reprograms pre-trained LLMs for time series forecasting without retraining by transforming numerical data into textual prompts. Similarly, [29] demonstrated that models such as GPT-3 and LLaMA-2 can forecast time series in a zero-shot setting without prior domain-specific training. [30] proposed LLM4TS, which aligns the LLM with the time series using a two-stage fine-tuning strategy. More recent work incorporates multimodal or external knowledge. [12] introduced a retrieval-augmented LLM for financial forecasting that combines time series and financial indicators, while [31] integrates numerical and textual data for multimodal prediction. Similarly, [6] explores explainable LLM-based forecasting by combining time series, metadata, and news. Despite these advances, most methods include covariates in a fixed way without systematically exploring which combinations are most effective. Existing approaches, such as PromptCast proposed by [11], which adapts prompt-based learning to structured time series tasks, and the method introduced by [5], which investigates LLM performance across different data patterns and proposes prompt engineering strategies to improve generalization, can both be accommodated within our framework.

Table 1 provides a summary of the use of LLMs for time series forecasting. As the table shows, our study is unique in designing a set of custom prompts to incorporate covariates such as time embeddings into the LLM input, while also conducting a systematic evaluation of different covariate combinations to identify the most effective configuration to improve predictive accuracy.

Table 1. Summary of recent studies on time series forecasting using LLMs.

Paper LLM(s) Used Contribution Dataset(s) External Knowledge Covariate Integration Prompt Selection Covariate Selection
[11] GPT-2, BART, T5 Proposed PromptCast, a prompt-based forecasting paradigm that leverages LLMs through natural language inputs to perform zero-shot and few-shot time series prediction. PISA (City Temp., Electricity, Visitor Flow) Yes Yes No No
[29] GPT-3, LLaMA-2 Introduced LLMTime: reframed time series forecasting as a next-token prediction task, demonstrating LLMs’ capabilities in zero-shot settings. 29 benchmark datasets No No No No
[2] GPT-2, BERT, LLaMA-7B Proposed Time-LLM: reprograms LLMs via prompt templates, achieving effective forecasting without altering model weights. ETT, M4 No No No No
[30] GPT-2 Presented LLM4TS, a two-stage fine-tuning framework to align pre-trained LLMs with time series forecasting tasks with minimal data. Seven real-world datasets No No No No
[5] GPT-3.5-turbo, GPT-4-turbo, Gemini-Pro, LLaMA-2-13B Analyzed LLMs’ strengths in trend/seasonal data and proposed methods to enhance forecasting via prompt engineering and contextual augmentation. Darts (Air Passengers, ETT, etc.) Yes No Yes No
[12] StockLLM (fine-tuned LLaMA3.2-1B) Developed a RAG framework for financial forecasting; includes FinSeer, an LLM-guided retriever to extract relevant historical sequences. Stock prices + 20 financial indicators Yes No No No
[6] GPT-4, Open LLaMA Demonstrated that LLMs can generate explainable forecasts by integrating multi-modal financial inputs, including time series, metadata, and news. NASDAQ-100 (prices, metadata, financial news) Yes No No No
[32] GPT-2 Proposed TimeCMA, an LLM-empowered framework for multivariate time series forecasting that aligns time series and prompt embeddings via cross-modality alignment. 8 real-world multivariate time series datasets No No No No
[33] GPT-2 Proposed S2IP-LLM, a framework that aligns time series embeddings with LLM semantic space via cross-modality alignment, using semantic anchors as prompts. Multiple benchmark datasets No No No No
[31] GPT-2 Proposed GPT4MTS, a prompt-based LLM framework integrating numerical and textual data for multimodal time series forecasting. GDELT-based multimodal time series dataset Yes No No No
This Study GPT-4o-mini Designed a novel prompting strategy that explicitly encodes external covariates (e.g., time embeddings) into LLM inputs, and systematically investigates which combinations yield the best forecasting performance. Evaluates across multiple forecast horizons, datasets, and prompt structures. WHO FluNet, Oakland Call Center, Air Passengers Yes Explicit (horizon-aligned) Validation Validation

3. Proposed framework

In the time series forecasting context, three quantities define the context: the observed data, which consists of historical target values; the forecast horizon, which specifies how many future time steps are to be predicted; and the covariates, which are known exogenous variables aligned with the time series that may influence future outcomes and are available both for the observed data and the forecast horizon. Our objective is to predict future values yT+1,...,yT+H given a history of past observations y1,...,yT and the covariates 𝐱1,...,𝐱T+H where T and H denote the length of the observed data and the forecast horizon, respectively. To assess model performance, we partition the dataset into a validation set and a test set, where the validation set is used to choose the best prompt-covariate pair. A key step is the construction of prompts that encode these three core components into natural language.

3.1. Prompt structures for forecasting

Our objective is to design a set of prompts that are deliberately simple. Specifically, we focus on prompts that rely solely on observed time series data and avoid the use of external knowledge or example completions. Each prompt should include an instruction or a natural language description of the forecasting task, together with encoded data that contains previous observations and covariates in a structured format. Then it should conclude with a clear forecast request that directs the model to generate predictions for the specified forecast horizon.

Although there are infinitely many ways to construct prompts for time series forecasting using LLMs, ranging from raw sequences to detailed explanatory formats, we categorize prompts based on how they organize and present target values and covariates to the model and consider three representative categories:

  • Coupled Prompt. Each observation is represented as a key-value pair, combining covariates with their corresponding target values in an ordered sequence. Forecasting is performed by adding future covariates without associated targets. This format leverages the autoregressive nature of LLMs by mimicking the standard language modeling task, where the model learns to associate sequences of covariates and targets in a contiguous stream. It is simple to implement and aligns well with token-based sequence modeling, which makes it a possible fit for LLMs.

  • Decoupled Prompt. Covariates and target values are separated into distinct lists. The prompt contains both target values and covariates for the observed data and asks to forecast future targets using a list of upcoming covariates. This separation may allow the LLM to differentiate between the input features and the forecasting target.

  • Contextualized Prompt. This format extends the Decoupled structure by adding context about the covariates. The goal is to help the LLM recognize recurring patterns (e.g., seasonal or weekly cycles) by emphasizing their importance. The prompt avoids revealing dataset-specific details and instead leverages the pre-trained knowledge of the LLM through high-level information.

As a benchmark, we include the No-Covariate Prompt, which omits exogenous information entirely and two prompting strategies inspired by previous work:

  • No-Covariate prompt (Baseline). Only the raw sequence of target values without any accompanying temporal or contextual information is included, which serves as a baseline to evaluate the effect of including covariates.

  • PromptCast [11]. This format presents the input as a compressed natural language sentence that implicitly integrates covariate and target information. It avoids rigid structural alignment between covariates and target values and instead relies on the ability of the model to interpret the sentence holistically.

  • Knowledge-Guided prompt [5]. This format builds on the Decoupled structure by including domain-specific knowledge, such as the subject or source of the time series (e.g., ’This series represents US electricity demand’).

The prompt templates corresponding to each of the prompting strategies are provided in Table 2.

Table 2. Prompt formats used in the study.

Prompt Type Prompt Template
No–Covariate data: [y1, y2, ..., yT]. Predict the next H values of the time series. Just return the values as a list. No explanation.
Coupled 𝐱1:y1, 𝐱2:y2, ... 𝐱T:yT, 𝐱T+1:, 𝐱T+2:, ... 𝐱T+H: Predict the next H values in the time series. Just return the values as a list. No explanation.
Decoupled Data: [y1,y2,y3,...,yT]. Covariates: [𝐱1,𝐱2,𝐱3,...,𝐱T]. Prediction covariates: [𝐱T+1,𝐱T+2,...,𝐱T+H]. Predict the next H values of the time series. Just return the prediction values as a list. No explanation.
Contextualized Data:[y1,y2,y3,...,yT]. Covariates:[𝐱1,𝐱2,𝐱3,...,𝐱T]. Prediction covariates: [𝐱T+1,𝐱T+2,...,𝐱T+H]. The sequence represents a univariate time series with aligned covariates. These covariates exhibit recurring patterns (e.g., weekly or seasonal cycles) that influence the behavior of the series. Use both the observed values and the structure of the covariates to identify trends. Predict the next H values based on the observed sequence and the upcoming covariate pattern. Just return the prediction values as a list. No explanation.
PromptCast [11] From 𝐱1 to 𝐱T, there were [y1,y2,y3,...,yT] values recorded. Predict the next H values. Just return the prediction values as a list of numbers. No explanation.
Knowledge-Guided [5] Same structure as the Decoupled format, but includes additional dataset or domain-specific information to guide forecasting (e.g., “This time series represents energy demand in the U.S.”).

3.2. Prompt and covariate selection

Selecting which prompt format to use and which covariate to include is a critical design decision. While many real-world datasets contain multiple time-dependent covariates, not all contribute equally to the accuracy of the forecast. We adopt a validation-based approach to guide both prompt selection and covariate integration in a coordinated manner, and the prompt-covariate structure that produces the best forecast accuracy in a hold-out validation set is selected. Although this approach is computationally expensive, it can provide a reliable basis for selecting effective prompting designs.

Specifically, we embed each covariate in the prompt structure and assess its forecast performance on the validation set. This procedure allows us to evaluate both the standalone predictive value of each covariate and its compatibility with the chosen prompt. By jointly considering the prompt structure and covariate effectiveness, we ensure that our final design maximizes forecasting accuracy. The prompt–covariate pair that results in the best validation performance is then chosen. Incorporating covariates can change the tokens given to the model and may make the forecast query more similar to previous cases with the same covariate values. The model’s attention can then focus on matching historical segments and may act like a similarity-weighted average over past patterns that align with the upcoming covariate. Different attention heads may capture recurring structures such as weekly or yearly seasonality. In this case, informative and well-aligned covariates sharpen attention and improve accuracy, while uninformative or mismatched covariates weaken attention and reduce performance. However, incorporating auxiliary information (e.g., covariates) is not free and may introduce certain trade-offs.

Data leakage is a critical concern when using LLMs for forecasting. If the prompt contains domain-specific information, it may reveal information, and the model may produce overly optimistic forecasting performance that fails to generalize to unseen future data. However, added prompt complexity may also lead to failures. In general, this is an uncontrollable and unmeasurable error, and prompts such as [5] could be vulnerable to this risk.

In the context of LLM-based forecasting, the computational overhead of different prompt formats is generally similar, but the monetary cost is largely determined by the number of tokens processed during inference. Each prompt incurs a usage fee proportional to its token length, which includes instructions, time indices, covariate values, formatting symbols, and surrounding natural language. Since the main body of the prompt is used for presenting the time series, adding covariates approximately doubles the token count relative to simple baselines, thereby nearly doubling the cost. In large-scale deployments where thousands of forecasts may be generated daily, these differences translate into substantial financial and computational overhead. Thus, covariate integration should be prioritized only when simpler and shorter prompts do not deliver sufficient accuracy.

4. Numerical experiments

We evaluate our approach with OpenAI’s GPT-4o-mini model on three real-world datasets from healthcare, service operations, and transportation, each widely used in time series prediction tasks [29,34]. In all cases, covariates are extracted directly from the datasets, such as calendar date, month, or day-of-week, to capture temporal and seasonal patterns beyond the raw time series. The computational cost of the algorithm depends on the large language model being used. For GPT-4o-mini, the wall-clock time required to process the prompt and generate predictions is typically on the order of a few seconds, often less than five seconds. The monetary cost, however, varies across models and depends on the number of input and output tokens. For GPT-4o-mini, the API pricing is approximately $0.15 per million input tokens and $0.60 per million output tokens. For the first experiment, for instance, the total cost was less than $0.01.

The first dataset consists of weekly influenza positive cases obtained from the WHO FluNet dataset that spans the years 2016–2024. We focus on univariate time series corresponding to Influenza A weekly positive cases in the United States, and the covariates include the calendar year, the calendar month (e.g., January or February), and a combined year-week indicator (e.g., 2024-W01), which are supposed to help encode temporal and seasonal structure. For evaluation, we used 25 weeks for validation and 25 weeks for testing within the year 2024, and assessed performance over 1-, 2-, and 5-week forecast horizons.

The second dataset is the Oakland Call Center dataset, which contains the records of daily call arrival volumes from 2012 to 2014. We restrict attention to 30-minute interval call volumes between 8:00 AM and 5:00 PM, and for evaluation, we select a five-week interval with three weeks for validation and two weeks for testing. The forecasting horizons include short-term (one day ahead) and mid-term (one week ahead) predictions, and the covariates are the exact date and the day of the week.

The third dataset is the Air Passengers time series, which reports monthly totals of international airline passengers from 1949 to 1960 (144 observations). We used the last four years (January 1956 to December 1959), with the first two years reserved for validation and the final two years for testing. The forecast horizons are 1, 6, and 12 months, and the covariates consist of the exact date and a categorical month-of-year indicator.

Although large pretrained models may contain general knowledge about calendar structure, this knowledge is not automatically activated in a forecasting prompt unless the temporal role of the covariates is made explicit. In particular, the model does not inherently know that a sequence of values such as years, months, dates, or days of the week should be interpreted as calendar-based covariates that are aligned with the target observations and should be used to infer recurring temporal patterns. Without such explicit alignment, these entries may simply be treated as additional tokens or labels rather than as structured temporal information relevant to forecasting. Therefore, the contribution of our prompt design is not merely to remind the model of calendar facts, but to organize the historical target values and future prediction points in a way that makes their temporal correspondence visible to the model. In this sense, the covariates serve as structured anchors that connect past observations to future forecast points. Nevertheless, we do not disentangle whether the improvement comes from the semantic content of the calendar variables, the explicit prompt structure, or the interaction between the two. Comparing explicit calendar covariates with implicit temporal cues, such as week-of-year descriptions, represents an important direction for future work.

To evaluate predictive accuracy, this article reports three commonly used metrics: Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Root Mean Squared Error (RMSE) [35]. The MAE measures the average magnitude of prediction errors without considering their direction:

MAE=1n∑i=1n|y^i−yi|,

where y^i is the predicted value and yi is the ground truth at time point i. The MAPE expresses this error as a percentage of the actual values, providing a scale-independent measure of accuracy:

MAPE=1n∑i=1n|y^i−yiyi|.

While MAPE is widely used, we also calculate RMSE, which penalizes larger errors more heavily and preserves the original units of the target variable.

RMSE=1n∑i=1n(y^i−yi)2.

Together, these metrics provide a comprehensive view of model performance.

4.1. Prompt–covariate pair selection

As this step represents the first attempt to systematically leverage covariates for time series forecasting, a validation-based approach is adopted to select the best prompt–covariate pair for each dataset. Table 3 reports the validation and test results across the datasets for three prompt designs: Coupled, Decoupled, and Contextualized.

Table 3. Forecasting results across datasets by horizon, covariate, and prompt design for validation and test sets.

Dataset Forecast Horizon Covariate Prompt Validation Test
RMSE MAE MAPE (%) RMSE MAE MAPE (%)
Influenza 1 Year Coupled 564.18 315.16 8.29 1964.97 797.32 15.11
Decoupled 12841.83 4985.04 476.44 17494.99 9051.56 1615.06
Contextualized 4352.89 3233.92 288.48 12008.47 5343.28 1071.19
Month Coupled 768.26 505.08 10.38 1682.17 592.28 13.78
Decoupled 10888.62 5414.68 377.94 16262.22 8857.16 1222.88
Contextualized 15614.99 8496.88 628.22 13111.15 8963.92 1470.28
Date Coupled 1084.00 556.48 9.52 1343.32 540.28 13.70
Decoupled 2912.15 1729.63 151.87 9107.15 3010.60 460.45
Contextualized 23521.77 14234.68 1681.51 32298.58 23936.16 4188.99
2 Year Coupled 1179.99 635.04 12.92 1969.99 808.36 17.84
Decoupled 2919.09 2140.24 121.35 3463.82 1806.20 150.00
Contextualized 7990.04 4234.16 429.99 17626.56 9044.36 2053.65
Month Coupled 738.45 438.24 9.52 1724.58 1238.56 30.36
Decoupled 16424.49 7586.64 711.53 7239.28 4792.96 503.03
Contextualized 4232.21 3146.80 176.02 8127.52 7007.72 1115.88
Date Coupled 1075.28 603.72 10.00 2976.31 757.52 18.72
Decoupled 23783.87 13662.32 1867.49 12413.28 4586.12 563.93
Contextualized 22495.99 10966.08 1433.17 34314.72 23307.08 3630.98
5 Year Coupled 1193.08 789.48 18.47 2250.02 1573.40 32.27
Decoupled 4740.10 3638.40 286.19 6014.52 2895.80 95.27
Contextualized 6793.82 5111.72 447.85 6720.57 5233.00 751.63
Month Coupled 1374.72 792.20 13.85 3805.09 1197.32 39.67
Decoupled 3975.82 3029.48 267.34 14610.38 9575.44 1070.37
Contextualized 6835.68 5371.36 525.50 8875.65 7188.44 946.63
Date Coupled 1407.35 819.56 18.86 6190.20 2579.56 41.41
Decoupled 5045.80 3490.08 89.22 37948.85 21213.24 1099.31
Contextualized 19709.42 15839.20 1645.34 50079.06 45460.72 8482.24
Call Center 1 Date Coupled 155.62 116.19 33.38 152.63 107.50 29.66
Decoupled 175.54 130.48 40.07 166.90 115.43 33.21
Contextualized 173.12 127.38 39.41 167.26 122.54 37.53
Day of Week Coupled 115.97 75.76 17.04 91.19 56.18 14.69
Decoupled 193.90 144.90 42.79 192.50 147.29 42.83
Contextualized 179.74 136.14 41.22 177.56 133.89 39.38
7 Date Coupled 196.18 157.05 42.51 178.10 128.54 35.90
Decoupled 187.66 135.52 42.26 178.26 131.25 40.25
Contextualized 235.56 175.43 53.71 177.71 130.96 40.09
Day of Week Coupled 109.65 72.24 15.40 86.56 51.64 12.93
Decoupled 209.13 151.62 47.27 200.39 152.71 45.53
Contextualized 213.20 153.38 47.63 190.65 136.57 42.41
Air Passengers 1 Date Coupled 39.22 30.21 8.39 41.47 31.25 6.78
Decoupled 88.25 54.38 15.02 84.56 62.96 14.58
Contextualized 77.91 52.92 14.40 88.94 68.13 16.11
Month Coupled 50.26 37.21 10.48 43.74 32.08 7.34
Decoupled 46.74 37.92 10.04 52.72 43.71 9.71
Contextualized 48.09 37.83 10.03 55.93 46.88 10.39
6 Date Coupled 47.45 41.88 11.50 55.95 45.33 9.75
Decoupled 126.00 95.79 26.54 88.62 73.13 16.00
Contextualized 75.85 62.46 16.88 94.42 68.04 14.30
Month Coupled 29.47 26.92 7.42 70.90 56.50 12.72
Decoupled 81.07 70.79 19.82 68.85 54.42 12.39
Contextualized 59.67 46.42 11.77 48.83 35.42 7.43
12 Date Coupled 27.07 20.08 5.28 31.26 25.08 5.63
Decoupled 56.77 42.79 11.40 123.69 94.67 19.62
Contextualized 86.14 63.71 18.18 98.84 85.71 19.46
Month Coupled 19.52 14.96 3.86 33.84 30.54 6.68
Decoupled 76.00 59.67 16.33 73.18 57.79 12.98
Contextualized 60.67 50.50 13.18 82.75 68.54 15.13

Several key observations emerge. In about 85% of the cases, the best prompt in the validation set also delivers the best performance on the test set. Moreover, the best prompt–covariate pairs identified in validation either remain the best or achieve accuracy comparable to the best on the test set, depending on the chosen metric. For instance, in the Call Center dataset, the Coupled prompt with Day of Week as a covariate achieves the best results in both validation and test sets. In the Influenza dataset with a forecast horizon of 2, the Coupled prompt with the Month covariate achieves the best performance across all metrics in validation, and on the test set, this pair yields the lowest RMSE, but not the lowest MAE or MAPE.

We emphasize that this explanation is based on prompt structure and observed performance, not on a direct analysis of attention weights or internal model states. Therefore, we do not claim that the Coupled prompt is theoretically optimal or that its advantage will hold for all LLMs, datasets, or covariate types. A more detailed ablation study that varies prompt length, delimiters, and the degree of structural alignment, or directly analyzes model attention patterns, would be valuable future work for explaining why different prompt formats perform differently. We also observe that the Coupled prompt achieves strong and stable performance across datasets, generally outperforming both Decoupled and Contextualized alternatives by large margins. Another observation is that sometimes the optimal covariate depends on the forecasting horizon. For example, when predicting two weeks ahead in the Influenza dataset, the Month covariate yields the best results, whereas for a five-week forecast horizon, the Year covariate performs better. Taken together, these results confirm that the validation set provides a reliable guide for selecting prompt–covariate pairs.

4.2. Comparison with no-covariate baseline

We next compared the best-performing prompt with the results of the No-Covariate prompt, which represents the simplest form of time series forecasting with LLMs. Table 4 reports the results across all datasets and forecast horizons and Fig 2 depicts the forecast sample paths. For the Influenza dataset, incorporating a covariate reduces RMSE by 76%, 53%, and 73% at forecast horizons of 1, 2, and 5 weeks, respectively. In the Call Center dataset, covariates lead to error reductions of 47% for the 1-day forecast horizon and 54% for the 7-day forecast horizon. For Air Passengers, the gains are equally substantial, with decreases of 54%, 42%, and 70% over forecast horizons of 1, 6, and 12 months, respectively. Although covariate integration consistently improves upon the No-Covariate prompt, the magnitude of the improvement depends on the chosen covariate. For example, in the Call Center dataset with a 7-day horizon, selecting Date as the covariate produces performance close to the No-Covariate prompt, which highlights the importance of careful covariate selection.

Table 4. Forecasting results of the coupled prompt versus the no-covariate prompt across datasets, forecast horizons, and covariates for both validation and test sets.

Dataset Forecast Horizon Covariate Validation Test
RMSE MAE MAPE (%) RMSE MAE MAPE (%)
Influenza 1 No–Covariate 1385.61 859.28 23.10 5673.10 1703.80 44.82
Year 564.18 315.16 8.29 1964.97 797.32 15.11
Month 768.26 505.08 10.38 1682.17 592.28 13.78
Date 1084.00 556.48 9.52 1343.32 540.28 13.70
2 No–Covariate 9822.05 3546.88 128.86 3404.30 1269.20 23.22
Year 1180.00 635.04 12.92 1969.99 808.36 17.84
Month 738.45 438.24 9.52 1724.58 1238.56 30.36
Date 1075.28 603.72 10.00 2976.31 757.52 18.72
5 No–Covariate 7057.82 4350.56 520.75 8530.29 3411.64 78.25
Year 1193.08 789.48 18.47 2250.02 1573.40 32.27
Month 1374.72 792.20 13.85 3805.09 1197.32 39.67
Date 1407.35 819.56 18.86 6190.20 2579.56 41.41
Call Center 1 No–Covariate 173.12 124.67 38.85 172.54 132.93 39.40
Date 155.62 116.19 33.38 152.63 107.50 29.66
Day of Week 115.97 75.76 17.04 91.19 56.18 14.69
7 No–Covariate 148.03 108.14 31.49 187.63 144.29 43.08
Date 196.18 157.05 42.51 178.10 128.54 35.90
Day of Week 109.65 72.24 15.40 86.56 51.64 12.93
Air Passengers 1 No–Covariate 64.09 50.33 14.12 91.65 63.54 14.87
Date 39.22 30.21 8.39 41.47 31.25 6.78
Month 50.26 37.21 10.48 43.74 32.08 7.34
6 No–Covariate 80.24 66.42 18.33 96.93 79.38 18.10
Date 47.45 41.88 11.50 55.95 45.33 9.75
Month 29.47 26.92 7.42 70.90 56.50 12.72
12 No–Covariate 85.12 66.21 18.81 106.84 77.83 17.52
Date 27.07 20.08 5.28 31.26 25.08 5.63
Month 19.52 14.96 3.86 33.84 30.54 6.68

Fig 2. Forecasts of the Coupled prompt with covariate integration versus the No-Covariate prompt across datasets, forecast horizons, and covariates for both validation and test sets.

Fig 2

4.3. Comparison with existing prompting strategies

To assess the comparative effectiveness of our approach, we benchmark the Coupled prompt against PromptCast and the Knowledge-Guided prompt, with the results summarized in Table 5 and the forecasts shown in Fig 3. The results show that the Coupled prompt consistently matches or outperforms existing methods and offers both a lower average error and greater stability.

Table 5. Forecasting results using the Coupled prompt compared with PromptCast and the Knowledge-Guided prompt across datasets and forecast horizons.

Dataset Forecast Horizon Prompt Validation Test
RMSE MAE MAPE (%) RMSE MAE MAPE (%)
Influenza 1 Coupled 1084.00 556.48 9.52 1343.32 540.28 13.70
PromptCast 1866.47 872.52 57.91 2685.95 1073.36 18.12
Knowledge-Guided 2767.88 1731.24 140.50 5486.29 2358.44 249.49
2 Coupled 738.45 438.24 9.52 1724.58 1238.56 30.36
PromptCast 5217.58 2148.36 108.55 3282.67 1384.28 26.70
Knowledge-Guided 11645.30 5123.48 226.25 3566.82 1829.88 63.00
5 Coupled 1374.72 792.20 13.85 3805.09 1197.32 39.67
PromptCast 4198.49 2398.32 70.41 5223.67 2237.40 48.57
Knowledge-Guided 32420.12 15613.40 2448.87 6017.91 2443.12 42.23
Call Center 1 Coupled 115.97 75.76 17.04 91.19 56.18 14.69
PromptCast 152.50 110.29 34.19 173.95 129.89 39.22
Knowledge-Guided 154.76 116.29 35.35 202.57 154.68 40.05
7 Coupled 109.65 72.24 15.40 86.56 51.64 12.93
PromptCast 175.20 127.57 38.19 178.25 135.43 40.62
Knowledge-Guided 191.45 137.71 40.10 171.42 127.36 38.60
Air Passengers 1 Coupled 50.26 37.21 10.48 43.74 32.08 7.34
PromptCast 46.74 37.17 9.99 69.85 55.13 12.40
Knowledge-Guided 50.31 43.04 11.37 61.83 51.13 11.37
6 Coupled 29.47 26.92 7.42 70.90 56.50 12.72
PromptCast 56.80 47.29 13.15 90.87 74.71 16.71
Knowledge-Guided 73.92 60.08 16.27 59.69 48.33 10.06
12 Coupled 19.52 14.96 3.86 33.84 30.54 6.68
PromptCast 63.26 48.75 12.85 80.42 64.13 13.38
Knowledge-Guided 98.47 76.04 19.42 95.81 76.21 15.82

Fig 3. Forecasts of the Coupled prompt compared to PromptCast and Knowledge-Guided prompt across datasets and forecast horizons.

Fig 3

For the Influenza dataset, the Coupled prompt achieves substantially lower error across all forecast horizons. PromptCast provides modest gains in a few cases but lacks consistency, while Knowledge-Guided prompting often performs poorly, especially at longer horizons where its errors grow dramatically. For the Call Center dataset, both PromptCast and Knowledge-Guided lag behind, and the reductions in RMSE are similar across forecast horizons, averaging around 50%. For the Air Passengers dataset, the difference at the 1-month forecast horizon is negligible, but at longer forecast horizons, the Coupled prompt achieves clear gains, reducing RMSE by at least 16% at 6 months and 59% at 12 months. Overall, the Coupled prompt shows superior performance across datasets and forecast horizons and yields more accurate forecasts, with the advantage more evident in longer-term predictions.

We then performed an analysis to assess the robustness of our comparison. To obtain multiple predictions, we queried the model with the same prompt 50 times, recorded the results for each replication, and derived the metrics for each of the sample predictions. Table 6 reports p-values from pairwise t-tests between the best prompt–covariate design and three alternatives (No Covariate, PromptCast, Knowledge-Guided) across datasets and forecast horizons for all evaluation metrics. The results show that the values are uniformly small on the test sets, indicating that the improvements are statistically reliable. For the Call Center dataset across both 1- and 7-step horizons, the p-values are all small, which shows that our method decisively outperforms all comparators. In the Influenza dataset, most test set p-values fall below 10–4, with a few larger cases in the range 10−2−10−1. The validation p-values are sometimes larger (up to 10–1), but still provide clear evidence of improvement. For the Air Passengers dataset, most results again show very small p-values, with only a few exceptions at short horizons where values remain small (on the order of 10–3). These observations confirm earlier findings in this section and in Section 4.2.

Table 6. P-values from pairwise t-tests between the best prompt–covariate design and three alternatives (No-Covariate, PromptCast, and Knowledge-Guided) across datasets and forecast horizons.

Dataset Forecast Horizon Prompt Validation Test
P-value (RMSE) P-value (MAE) P-value (MAPE) P-value (RMSE) P-value (MAE) P-value (MAPE)
Influenza 1 No–Covariate ≤10−4 ≤10−4 ≤10−4 2.03E-02 ≤10−4 ≤10−4
PromptCast ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
Knowledge-Guided ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
2 No–Covariate 1.06E-02 ≤10−4 ≤10−4 2.31E-02 1.09E-02 1.37E-01
PromptCast 1.70E-02 1.50E-02 4.29E-02 1.11E-02 5.88E-03 8.08E-02
Knowledge-Guided ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
5 No–Covariate 1.31E-01 1.15E-01 ≤10−4 ≤10−4 ≤10−4 ≤10−4
PromptCast 3.39E-02 3.09E-02 8.52E-03 3.55E-01 2.99E-01 6.56E-02
Knowledge-Guided ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
Call Center 1 No–Covariate ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
PromptCast ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
Knowledge-Guided ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
7 No–Covariate ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
PromptCast ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
Knowledge-Guided ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
Air Passengers 1 No–Covariate ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
PromptCast 7.14E-03 2.23E-03 1.11E-02 2.88E-03 1.69E-03 6.74E-03
Knowledge-Guided ≤10−4 ≤10−4 ≤10−4 5.10E-04 ≤10−4 ≤10−4
6 No–Covariate ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
PromptCast ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
Knowledge-Guided ≤10−4 ≤10−4 ≤10−4 7.05E-03 3.97E-05 9.82E-05
12 No–Covariate ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
PromptCast ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4
Knowledge-Guided ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4 ≤10−4

4.4. Sensitivity to missing covariates

Next, we examine how sensitive the forecasts are by analyzing the effect of random censoring of covariates on the evaluation metrics. Table 7 reports the forecast results when covariates are subjected to random censoring at different ratios. As the censoring level increases from 0.1 to 0.9, errors generally increase across datasets and forecast horizons, although some fluctuations remain due to randomness in the censoring process. This behavior reflects the loss of signal content in covariates as more entries are masked.

Table 7. Forecasting results across datasets, forecast horizons, and censoring levels.

Dataset Forecast Horizon Censoring Level Validation Test
RMSE MAE MAPE (%) RMSE MAE MAPE (%)
Influenza 1 0.1 746.54 444.28 12.50 1367.23 557.80 17.25
0.3 745.66 443.16 9.74 646.01 362.16 17.09
0.5 944.42 536.64 10.29 3500.96 1196.60 48.25
0.7 1005.52 627.76 12.77 2441.39 869.64 18.74
0.9 1153.40 661.96 12.15 1724.29 733.04 15.42
2 0.1 869.61 524.56 10.93 1890.55 829.00 17.93
0.3 1022.43 689.40 15.50 1876.15 836.52 28.11
0.5 2569.13 1464.04 25.33 2899.08 1206.48 25.69
0.7 2769.17 1701.28 35.63 2763.10 1281.64 30.39
0.9 3072.41 1772.96 170.18 2333.31 1049.16 50.76
5 0.1 3125.41 2151.76 49.04 3265.41 1284.44 28.58
0.3 3458.45 2128.96 33.83 4681.76 1826.24 25.37
0.5 5532.22 4045.76 442.86 2997.66 1379.00 34.11
0.7 5461.74 4195.52 101.92 4769.56 2072.04 36.49
0.9 11420.23 7122.88 160.64 5007.93 1958.96 33.08
Call Center 1 0.1 75.29 53.48 11.61 142.09 85.86 21.10
0.3 153.38 111.52 27.74 180.24 127.36 24.08
0.5 144.32 93.62 24.79 172.35 124.29 34.55
0.7 161.47 115.88 22.91 211.63 147.54 31.38
0.9 185.16 133.72 25.44 227.92 169.37 34.61
7 0.1 79.38 54.33 11.87 112.45 76.79 17.01
0.3 120.89 74.71 13.92 163.49 117.29 23.35
0.5 173.12 135.57 35.05 229.78 185.64 48.91
0.7 181.68 134.39 24.21 236.52 184.11 34.34
0.9 203.75 153.84 28.62 269.47 204.22 38.19
Air Passengers 1 0.1 46.29 35.21 9.94 44.13 35.42 8.14
0.3 46.69 37.33 10.53 55.62 43.92 10.02
0.5 82.07 63.21 17.85 90.17 70.04 16.50
0.7 70.61 54.96 14.98 60.84 44.50 9.76
0.9 79.48 63.42 17.60 87.10 72.04 16.59
6 0.1 43.29 37.71 10.37 49.34 41.29 8.86
0.3 58.63 50.50 13.99 37.22 30.75 6.84
0.5 103.25 87.71 23.59 117.42 91.54 21.74
0.7 87.28 69.17 18.23 76.98 59.50 12.93
0.9 100.26 86.00 23.80 122.91 107.29 25.56
12 0.1 25.21 21.00 5.67 51.78 34.75 8.55
0.3 21.83 18.92 5.18 40.53 37.58 8.28
0.5 84.44 63.88 19.43 72.54 61.92 13.37
0.7 102.17 84.21 24.53 86.35 76.17 17.53
0.9 124.60 106.71 29.75 119.74 88.42 20.99

In the Influenza dataset, error growth is clear at higher forecast horizons. The validation and test RMSE more than double as the censoring level increases from 0.1 to 0.9. The performance of the Coupled prompt stays better than that of the No-Covariate prompt up to a censoring level of 0.9 for forecast horizons of 1 and 2, but falls behind at the censoring level of 0.9 for the forecast horizon of 5. In the Call Center dataset, degradation is also evident. At the forecast horizon of 1 with the censoring level of 0.9, the No-Covariate prompt performs better, and a similar result appears at the censoring level of 0.5 for the forecast horizon of 7. In the Air Passengers dataset, the error also increases as the level of censoring increases. At censoring levels of 0.5 and above, covariate integration no longer performs better than the No-Covariate prompt. In general, these results confirm that censored covariates reduce the accuracy of forecasts. The overall pattern shows the importance of careful covariate selection and reliable measurement to achieve stable and accurate LLM-based forecasts, while also showing that even partially observed covariates can still improve accuracy to some extent.

4.5. Comparison with existing time series forecasting methods

The previous sections focused on comparing different prompting strategies and covariate designs within the LLM-based framework. To place these results in context, it is also important to benchmark against established forecasting approaches. In particular, we compare our method with ARIMA, a widely used classical statistical model, and LSTM, a standard deep learning baseline. This comparison allows us to assess whether covariate-informed LLM prompting can serve as a credible alternative to traditional forecasting methods. Table 8 reports the results for ARIMA and LSTM, while Table 5 summarizes the performance of our prompting strategies. All experiments were carried out in the same setting, using identical datasets, forecast horizons, and validation/test splits. For the Influenza dataset, ARIMA performs poorly with very high errors (e.g., 1-step test RMSE above 6600 and MAPE above 200%), and LSTM achieves better accuracy (1-step test RMSE 2766, MAE 1730). In contrast, the Coupled prompt with covariates reduces the 1-step RMSE to 1343 and MAE to 540, cutting error by more than 50% compared to LSTM. At the longer horizon of 5 weeks, the Coupled prompt still achieves a test RMSE of 3805, far lower than ARIMA (7459) or LSTM (6571), showing that LLM prompting remains competitive in noisy, high-variability series.

Table 8. Forecasting results of LSTM and ARIMA across datasets and forecast horizons.

Dataset Step Model Validation Test
RMSE MAE MAPE (%) RMSE MAE MAPE (%)
Influenza 1 LSTM 4848.98 3701.46 93.04 2766.92 1730.30 106.10
ARIMA 8593.71 7473.96 742.43 6603.52 4027.81 238.06
2 LSTM 5312.51 3845.87 78.40 3494.66 1423.81 56.22
ARIMA 7356.63 6313.16 632.88 6828.26 3691.23 185.82
5 LSTM 5959.22 4471.88 160.35 6571.31 3584.02 109.81
ARIMA 5624.28 4872.94 424.92 7459.18 5197.14 391.68
Call Center 1 LSTM 66.46 50.53 11.08 78.85 49.17 12.05
ARIMA 72.69 58.89 14.04 82.27 60.20 14.68
7 LSTM 81.27 65.11 16.40 82.36 52.89 13.07
ARIMA 64.39 46.22 10.55 101.04 87.19 19.23
Air Passengers 1 LSTM 30.94 26.86 6.93 78.16 71.28 15.37
ARIMA 42.05 31.56 7.91 62.57 46.64 9.44
6 LSTM 44.97 40.06 10.27 73.54 67.20 14.59
ARIMA 38.92 27.62 6.80 59.56 45.78 9.39
12 LSTM 27.62 21.28 5.31 84.71 77.09 16.58
ARIMA 14.65 9.75 2.45 59.63 53.03 11.68

For the Call Center dataset, LSTM achieves a test MAE of 49 at the 1-step horizon, and ARIMA follows with 60. The Coupled prompt with day-of-week covariate improves this to 56 MAE, which is better than ARIMA but slightly above LSTM for short horizons. However, at the 7-step horizon, LSTM degrades to 82 RMSE and ARIMA exceeds 100, while the Coupled prompt maintains only 87 RMSE and 52 MAE, representing around 40% lower MAE than ARIMA and comparable performance to LSTM in validation and test. For the Air Passengers dataset, ARIMA is strong at long horizons: at 12 steps, it reaches a test RMSE of 59 and MAPE of 11.7, while LSTM deteriorates to RMSE 85 and MAPE 16.6. The Coupled prompt achieves RMSE 33 and MAE 30 at the same horizon, nearly halving the error relative to ARIMA and outperforming LSTM by a large margin. At shorter horizons, such as 1 step, the Coupled prompt again improves upon both classical models (test RMSE 43 vs. 78 for LSTM and 62 for ARIMA). Overall, these results show that LLM-based prompting strategies are generally superior to ARIMA and competitive with LSTM across datasets.

5. Conclusion

In this study, we analyzed the incorporation of covariates in time series forecasting with commonly used LLMs by proposing a set of candidate prompts and exploring key considerations and limitations. The proposed approach was implemented on three real-world datasets, and the results demonstrated that employing a well-structured prompting strategy, together with the inclusion of a suitable covariate from the dataset, could significantly enhance the accuracy and quality of forecasts.

Despite these promising results, several important challenges remain. First, many current LLM-based forecasting approaches lack formal mechanisms for uncertainty quantification and may struggle to generalize across diverse application domains. A natural next step is to extend the framework to probabilistic forecasting, prompting for calibrated quantiles, prediction intervals, or full predictive distributions, and evaluating them with proper scoring rules and empirical coverage. Moreover, most existing work is restricted to univariate time series, although multivariate forecasting has wide applicability in real-world scenarios. Generalizing the method to multivariate settings by conditioning on multiple synchronized series and cross-series covariates, inspired by approaches such as TimeCMA [32], is a promising direction.

Data Availability

All data and code required to reproduce the results in this study are publicly available in the GitHub repository at https://github.com/mohammadmgh78/Informed-Forecasting, with no access restrictions. The repository contains: (1) Jupyter notebooks implementing the full pipeline for each forecasting experiment described in the manuscript (influenza incidence, call-center arrival counts, and airline passenger counts), organized into data preparation, LLM-based forecasting, and result-analysis stages; (2) the raw and processed input time series used in each experiment; (3) cached outputs from all LLM forecasting queries, one file per experimental configuration, allowing the result-analysis notebooks to be rerun without repeating the API calls; and (4) the aggregated result tables and figures presented in the manuscript.

Funding Statement

The author(s) received no specific funding for this work.

References

  • 1.The Verge. ChatGPT Weekly Users Reach 300 Million in December 2024. 2024. [cited 2025 Jan 22]. https://www.theverge.com/2024/12/4/24313097/chatgpt-300-million-weekly-users
  • 2.Jin M, Wang S, Ma L, Chu Z, Zhang JY, Shi X, et al. Time-llm: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:231001728. 2023.
  • 3.Akhavan A, Jalali MS. Generative AI and simulation modeling: how should you (not) use large language models like ChatGPT. Syst Dyn Rev. 2024;40(3). doi: 10.1002/sdr.1773 [DOI] [Google Scholar]
  • 4.Vemprala SH, Bonatti R, Bucker A, Kapoor A. ChatGPT for robotics: Design principles and model abilities. IEEE Access. 2024;12:55682–96. doi: 10.1109/access.2024.3387941 [DOI] [Google Scholar]
  • 5.Tang H, Zhang C, Jin M, Yu Q, Wang Z, Jin X, et al. Time series forecasting with LLMs: Understanding and enhancing model capabilities. SIGKDD Explor Newsl. 2025;26(2):109–18. doi: 10.1145/3715073.3715083 [DOI] [Google Scholar]
  • 6.Yu X, Chen Z, Ling Y, Dong S, Liu Z, Lu Y. Temporal data meets LLM–explainable financial time series forecasting. arXiv preprint arXiv:230611025. 2023.
  • 7.Meem JA, Rashid MS, Hristidis V. Modeling the impact of out-of-schema questions in task-oriented dialog systems. Data Min Knowl Disc. 2024;38(4):2466–94. doi: 10.1007/s10618-024-01039-6 [DOI] [Google Scholar]
  • 8.Karimian MA, Zolbanin HM, Delen D. Explainable deep learning architecture for multivariate time series forecasting. Data Min Knowl Discov. 2024. doi: 10.1007/s10618-023-00984-y [DOI] [Google Scholar]
  • 9.Lin H, Chung HS-H, Lin C, Xie D, Deng Q, Lyu M, et al. Improved fault diagnosis capability in CHBMCs: Counter design for multiple OC switches via an E-SVM unit. IEEE Trans Power Electron. 2026;41(2):2358–76. doi: 10.1109/tpel.2025.3618228 [DOI] [Google Scholar]
  • 10.Meng X, Xie D, Lin H, Lin C, Ge X, Liu Z. Dissipativity-based multiport stability root-cause identification and mitigation for solid-state transformers. IEEE Trans Ind Electron. 2026;73(7):9850–62. doi: 10.1109/tie.2026.3658639 [DOI] [Google Scholar]
  • 11.Xue H, Salim FD. PromptCast: A new prompt-based learning paradigm for time series forecasting. IEEE Trans Knowl Data Eng. 2024;36(11):6851–64. doi: 10.1109/tkde.2023.3342137 [DOI] [Google Scholar]
  • 12.Xiao M, Jiang Z, Qian L, Chen Z, He Y, Xu Y, et al. Retrieval-augmented Large Language Models for Financial Time Series Forecasting. arXiv preprint arXiv:250205878. 2025.
  • 13.Sandhya Arora MK. Forecasting the future: A comprehensive review of time series prediction techniques. J Electr Syst. 2024;20(2s):575–86. doi: 10.52783/jes.1478 [DOI] [Google Scholar]
  • 14.Sezer OB, Gudelek MU, Ozbayoglu AM. Financial time series forecasting with deep learning: A systematic literature review: 2005–2019. Appl Soft Comput. 2020;90:106181. doi: 10.1016/j.asoc.2020.106181 [DOI] [Google Scholar]
  • 15.Dingli A, Fournier KS. Financial time series forecasting – A deep learning approach. Int J Mach Learn Comput. 2017;7(5):118. doi: 10.18178/ijmlc.2017.7.5.632 [DOI] [Google Scholar]
  • 16.Kaushik S, Choudhury A, Sheron PK, Dasgupta N, Natarajan S, Pickett LA, et al. AI in healthcare: Time-series forecasting using statistical, neural, and ensemble architectures. Front Big Data. 2020;3:4. doi: 10.3389/fdata.2020.00004 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Koivisto M, Das K, Guo F, Sørensen P, Nuño E, Cutululis N, et al. Using time series simulation tools for assessing the effects of variable renewable energy generation on power and energy systems. WIREs Energy Environ. 2018;8(3). doi: 10.1002/wene.329 [DOI] [Google Scholar]
  • 18.Aviv Y. A time-series framework for supply-chain inventory management. Oper Res. 2003;51(2):210–27. doi: 10.1287/opre.51.2.210.12780 [DOI] [Google Scholar]
  • 19.Ariyo AA, Adewumi AO, Ayo CK. Stock price prediction using the ARIMA model. In: 2014 UKSim-AMSS 16th international conference on computer modelling and simulation. IEEE; 2014. pp. 106–12. [Google Scholar]
  • 20.Fattah J, Ezzine L, Aman Z, El Moussami H, Lachhab A. Forecasting of demand using ARIMA model. Int J Eng Bus Manage. 2018;10. doi: 10.1177/1847979018808673 [DOI] [Google Scholar]
  • 21.Chen P, Yuan H, Shu X. Forecasting crime using the ARIMA model. In: 2008 fifth international conference on fuzzy systems and knowledge discovery. vol. 5. IEEE; 2008. pp. 627–30. [Google Scholar]
  • 22.Yildiz B, Bilbao JI, Sproul AB. A review and analysis of regression and machine learning models on commercial building electricity load forecasting. Renew Sustain Energy Rev. 2017;73:1104–22. doi: 10.1016/j.rser.2017.02.023 [DOI] [Google Scholar]
  • 23.Bianco V, Manca O, Nardini S. Electricity consumption forecasting in Italy using linear regression models. Energy. 2009;34(9):1413–21. doi: 10.1016/j.energy.2009.06.034 [DOI] [Google Scholar]
  • 24.Obata S, Cieszewski CJ, Lowe RC III, Bettinger P. Random forest regression model for estimation of the growing stock volumes in Georgia, USA, using dense landsat time series and FIA dataset. Remote Sens. 2021;13(2):218. doi: 10.3390/rs13020218 [DOI] [Google Scholar]
  • 25.Banskota A, Kayastha N, Falkowski MJ, Wulder MA, Froese RE, White JC. Forest monitoring using landsat time series data: A review. Can J Remote Sens. 2014;40(5):362–84. doi: 10.1080/07038992.2014.987376 [DOI] [Google Scholar]
  • 26.Zeroual A, Harrou F, Dairi A, Sun Y. Deep learning methods for forecasting COVID-19 time-Series data: A Comparative study. Chaos Solitons Fractals. 2020;140:110121. doi: 10.1016/j.chaos.2020.110121 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Liang W, Zhang Y, Wu Z, Lepp H, Ji W, Zhao X, et al. Mapping the increasing use of llms in scientific papers. arXiv preprint arXiv:240401268. 2024.
  • 28.IBM Research. What are Large Language Models?. 2024. [cited 2025 Jan 22]. Available from: https://www.ibm.com/think/topics/large-language-models
  • 29.Gruver N, Finzi M, Qiu S, Wilson AG. Large language models are zero-shot time series forecasters. In: Proceedings of the 37th International Conference on Neural Information Processing Systems. NIPS ’23. Red Hook, NY, USA: Curran Associates Inc.; 2023. [Google Scholar]
  • 30.Chang C, Wang W-Y, Peng W-C, Chen T-F. LLM4TS: Aligning pre-trained LLMs as data-efficient time-series forecasters. ACM Trans Intell Syst Technol. 2025;16(3):1–20. doi: 10.1145/3719207 [DOI] [Google Scholar]
  • 31.Jia F, Wang K, Zheng Y, Cao D, Liu Y. Gpt4mts: Prompt-based large language model for multimodal time-series forecasting. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38. 2024. pp. 23343–51.
  • 32.Liu C, Xu Q, Miao H, Yang S, Zhang L, Long C, et al. TimeCMA: Towards llm-empowered multivariate time series forecasting via cross-modality alignment. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39. 2025. pp. 18780–8.
  • 33.Pan Z, Jiang Y, Garg S, Schneider A, Nevmyvaka Y, Song D. S2 IP-LLM: Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting. In: Forty-first International Conference on Machine Learning; 2024.
  • 34.Zheng Y, Zheng Z, Zhu T. A doubly stochastic simulator with applications in arrivals modeling and simulation. Oper Res. 2025;73(2):910–26. doi: 10.1287/opre.2021.0597 [DOI] [Google Scholar]
  • 35.Hyndman RJ, Koehler AB. Another look at measures of forecast accuracy. Int J Forecast. 2006;22(4):679–88. doi: 10.1016/j.ijforecast.2006.03.001 [DOI] [Google Scholar]

Decision Letter 0

Dinesh Nishad

18 Apr 2026

Dear Dr. Ghasemloo,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please submit your revised manuscript by Jun 02 2026 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

  • A letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

  • A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

  • An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

As the corresponding author, your ORCID iD is verified in the submission system and will appear in the published article. PLOS supports the use of ORCID, and we encourage all coauthors to register for an ORCID iD and use it as well. Please encourage your coauthors to verify their ORCID iD within the submission system before final acceptance, as unverified ORCID iDs will not appear in the published article. Only  the individual author can complete the verification step; PLOS staff cannot  verify ORCID iDs on behalf of authors.

We look forward to receiving your revised manuscript.

Kind regards,

Dinesh Kumar Nishad, PhD

Academic Editor

PLOS One

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1.Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at

https://journals.plos.org/plosone/s/file?id=wjVg/PLOSOne_formatting_sample_main_body.pdf and

https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. PLOS requires an ORCID iD for the corresponding author in Editorial Manager on papers submitted after December 6th, 2016. Please ensure that you have an ORCID iD and that it is validated in Editorial Manager. To do this, go to ‘Update my Information’ (in the upper left-hand corner of the main menu), and click on the Fetch/Validate link next to the ORCID field. This will take you to the ORCID site and allow you to create a new iD or authenticate a pre-existing iD in Editorial Manager.

3. Please note that PLOS One has specific guidelines on code sharing for submissions in which author-generated code underpins the findings in the manuscript. In these cases, we expect all author-generated code to be made available without restrictions upon publication of the work. Please review our guidelines at https://journals.plos.org/plosone/s/materials-and-software-sharing#loc-sharing-code and ensure that your code is shared in a way that follows best practice and facilitates reproducibility and reuse.

4. Please note that your Data Availability Statement is currently missing the repository name and/or the DOI/accession number of each dataset OR a direct link to access each database. If your manuscript is accepted for publication, you will be asked to provide these details on a very short timeline. We therefore suggest that you provide this information now, though we will not hold up the peer review process if you are unable.

5. If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise.

[Note: HTML markup is below. Please do not edit.]

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

Reviewer #1: Yes

Reviewer #2: Partly

Reviewer #3: Yes

**********

2. Has the statistical analysis been performed appropriately and rigorously? -->?>

Reviewer #1: Yes

Reviewer #2: N/A

Reviewer #3: N/A

**********

3. Have the authors made all data underlying the findings in their manuscript fully available??>

The PLOS Data policy

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English??>

Reviewer #1: Yes

Reviewer #2: Yes

Reviewer #3: Yes

**********

Reviewer #1: This manuscript presents a well-designed empirical study on enhancing large language models (LLMs) for time series forecasting via time-dependent covariates. The experimental framework is rigorous, utilizing diverse real-world datasets, adequate baseline comparisons, and robust statistical validations, making the methodology technically sound. However, a few critical issues regarding reference accuracy must be addressed prior to publication.

1. Reference 9: The citation for the "PromptCast" paper is incomplete and lists an incorrect year. Please update it to the official publication.

2. References 27 and 33: These entries are duplicates of the same paper by Gruver et al. Please remove the redundancy and consolidate the in-text citations.

3. References 30 and 34: Reference 30 cites the arXiv preprint of TimeCMA, while Reference 34 cites the officially accepted AAAI 2025 version. Please remove Reference 30 and direct all in-text citations to the published conference version (Reference 34).

Reviewer #2: This study investigates how time-dependent covariates can be integrated into prompt design for large language models to improve time series forecasting performance across multiple real-world datasets and forecasting settings. The manuscript needs a major revision to be accepted.

1. The abstract clearly states the conceptual goal, but it does not identify which LLMs, baseline forecasting methods, or evaluation metrics were used; including this information would make the reported improvements more interpretable and technically credible.

2. Although the abstract claims statistically significant improvements, it does not provide any representative quantitative results or effect sizes; adding concise numerical evidence would strengthen the impact of the contribution.

3. The introduction should clearly conclude with a distinct section highlighting the novel contributions of your work.

4. At the ending of the intro, it is advised to add a para that mentions briefly what each next section contains.

5. The literature review should benefit from more explorations of previous studies.

6. The discussion section needs to be expanded to more thoroughly analyze the results.

7. The first paragraph of the conclusion should succinctly summarize the contributions of the study in past tense.

8. The second paragraph of the conclusion should provide clear and actionable future recommendations.

9. Equations are not properly cited, please add original references.

10. Merge the related sections please.

Reviewer #3: This paper proposes a prompt engineering pipeline for LLM-based time series forecasting that incorporates time-dependent covariates, evaluated on three public datasets using GPT-4o-mini.

1. The entire study relies on a single proprietary model, GPT-4o-mini. It is unclear whether the observed benefits of covariate integration are specific to GPT-4o-mini's pretraining data and tokenization or whether they hold for other LLMs such as open-source models (LLaMA-3, Mistral, or Gemma). Without at least one additional model, the results cannot be claimed as general findings about LLM-based forecasting; they describe GPT-4o-mini behavior on these three datasets.

2. The "Coupled" prompt is identified as the best format across nearly all settings, yet the paper offers no theoretical or mechanistic explanation for why this is the case. The attention-based intuition offered in Section 4.2 is speculative and not supported by any attention analysis or ablation. A brief experiment examining why decoupled and contextualized formats underperform, for example by varying prompt length or testing with and without structural delimiters, would strengthen this central claim.

3. The validation-based selection procedure requires running inference for every (prompt format, covariate) pair on the validation set. With 3 prompt types and 3 covariate options per dataset, this is 9 API calls per prediction step, and the paper acknowledges that covariate inclusion roughly doubles token count. The reported cost in dollars and wall-clock time is absent. Given that PLOS ONE serves a broad audience including practitioners without large API budgets, a cost table or at least an order-of-magnitude estimate is needed.

4. All three covariates used, year, month, date, and day-of-week, are calendar attributes that are trivially available in any time series. The paper claims to focus on "covariates that are not tied to a specific dataset," but this design choice also limits the contribution. Calendar features are already internally represented in large pretrained models, so it is not clear whether the gains come from the prompt engineering format or simply from reminding the model of a date it could infer from context. An experiment comparing explicit date covariates versus implicit cues (e.g., telling the model "this is week 48 of the year" versus providing the date string) would clarify this.

5. The following references are recommended to consider to strengthen the manuscript's discussion.

[1] https://doi.org/10.1109/TPEL.2025.3618228

[2] https://doi.org/10.1109/TIE.2026.3658639

**********

what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review?  For information about this choice, including consent withdrawal, please see our Privacy Policy

Reviewer #1: No

Reviewer #2: No

Reviewer #3: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

To ensure your figures meet our technical requirements, please review our figure guidelines: https://journals.plos.org/plosone/s/figures

You may also use PLOS’s free figure tool, NAAS, to help you prepare publication quality figures: https://journals.plos.org/plosone/s/figures#loc-tools-for-figure-preparation.

NAAS will assess whether your figures meet our technical requirements by comparing each figure against our figure specifications.

PLoS One. 2026 Oct 1;21(10):e0358970. doi: 10.1371/journal.pone.0358970.r002

Author response to Decision Letter 1


27 Jun 2026

We thank the editor and reviewers for their helpful and constructive comments. We have made every effort to address the comments carefully and believe that the revisions have significantly strengthened the manuscript. For detailed point-by-point responses, please refer to the accompanying Response to Reviewers file. If any additional revisions or clarifications are needed, we would be happy to provide them.

Attachment

Submitted filename: Response to Reviewers.pdf

pone.0358970.s002.pdf (103KB, pdf)

Decision Letter 1

Dinesh Nishad

18 Aug 2026

Dear Dr. Ghasemloo,

Thank you for submitting your manuscript to PLOS One. After careful consideration, we feel that it has merit but does not fully meet PLOS One’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please submit your revised manuscript by Oct 02 2026 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

  • A letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

  • A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

  • An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: https://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS One offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

As the corresponding author, your ORCID iD is verified in the submission system and will appear in the published article. PLOS supports the use of ORCID, and we encourage all coauthors to register for an ORCID iD and use it as well. Please encourage your coauthors to verify their ORCID iD within the submission system before final acceptance, as unverified ORCID iDs will not appear in the published article. Only  the individual author can complete the verification step; PLOS staff cannot  verify ORCID iDs on behalf of authors.

We look forward to receiving your revised manuscript.

Kind regards,

Dinesh Kumar Nishad, PhD

Academic Editor

PLOS One

Journal Requirements:

If the reviewer comments include a recommendation to cite specific previously published works, please review and evaluate these publications to determine whether they are relevant and should be cited. There is no requirement to cite these works unless the editor has indicated otherwise.

Please review your reference list to ensure that it is complete and correct. If you have cited papers that have been retracted, please include the rationale for doing so in the manuscript text, or remove these references and replace them with relevant current references. Any changes to the reference list should be mentioned in the rebuttal letter that accompanies your revised manuscript. If you need to cite a retracted article, indicate the article’s retracted status in the References list and also include a citation and full reference for the retraction notice.

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

Reviewer #2: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions??>

Reviewer #2: (No Response)

**********

3. Has the statistical analysis been performed appropriately and rigorously? -->?>

Reviewer #2: (No Response)

**********

4. Have the authors made all data underlying the findings in their manuscript fully available??>

The PLOS Data policy

Reviewer #2: (No Response)

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English??>

Reviewer #2: (No Response)

**********

Reviewer #2: Well responded, now kindly conduct minor English proofreading for the manuscript as there are some typos.

**********

what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review?  For information about this choice, including consent withdrawal, please see our Privacy Policy

Reviewer #2: No

**********

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

To ensure your figures meet our technical requirements, please review our figure guidelines: https://journals.plos.org/plosone/s/figures

You may also use PLOS’s free figure tool, NAAS, to help you prepare publication quality figures: https://journals.plos.org/plosone/s/figures#loc-tools-for-figure-preparation.

NAAS will assess whether your figures meet our technical requirements by comparing each figure against our figure specifications.

PLoS One. 2026 Oct 1;21(10):e0358970. doi: 10.1371/journal.pone.0358970.r004

Author response to Decision Letter 2


26 Aug 2026

We sincerely thank the Academic Editor and Reviewer 2 for their careful evaluation of our revised manuscript. We are grateful that Reviewer 2 confirmed that all comments from the previous round had been addressed. We also appreciate the request for a final minor English proofreading pass.

In response, we carefully proofread the entire manuscript and made several minor corrections throughout the paper.

All changes made in this revision are highlighted in blue in the marked-up manuscript. These revisions are editorial in nature and do not alter the methodology, experimental results, statistical analysis, or scientific conclusions of the paper.

Attachment

Submitted filename: Response to Reviews Minor Revision.pdf

pone.0358970.s003.pdf (71.2KB, pdf)

Decision Letter 2

Dinesh Nishad

8 Sep 2026

Informed forecasting: Leveraging auxiliary knowledge to boost large language model performance on time series forecasting

PONE-D-26-08442R2

Dear Dr. Ghasemloo,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice will be generated when your article is formally accepted. Please note, if your institution has a publishing partnership with PLOS and your article meets the relevant criteria, all or part of your publication costs will be covered. Please make sure your user information is up-to-date by logging into Editorial Manager at Editorial Manager® and clicking the ‘Update My Information' link at the top of the page. For questions related to billing, please contact billing support.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Dinesh Kumar Nishad, PhD

Academic Editor

PLOS One

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

Reviewer #2: (No Response)

**********

2. Is the manuscript technically sound, and do the data support the conclusions??>

Reviewer #2: (No Response)

**********

3. Has the statistical analysis been performed appropriately and rigorously? -->?>

Reviewer #2: (No Response)

**********

4. Have the authors made all data underlying the findings in their manuscript fully available??>

The PLOS Data policy

Reviewer #2: (No Response)

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English??>

Reviewer #2: (No Response)

**********

Reviewer #2: (No Response)

**********

what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review?  For information about this choice, including consent withdrawal, please see our Privacy Policy

Reviewer #2: No

**********

Acceptance letter

Dinesh Nishad

PONE-D-26-08442R2

PLOS One

Dear Dr. Ghasemloo,

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS One. Congratulations! Your manuscript is now being handed over to our production team.

At this stage, our production department will prepare your paper for publication. This includes ensuring the following:

* All references, tables, and figures are properly cited

* All relevant supporting information is included in the manuscript submission,

* There are no issues that prevent the paper from being properly typeset

You will receive further instructions from the production team, including instructions on how to review your proof when it is ready. Please keep in mind that we are working through a large volume of accepted articles, so please give us a few days to review your paper and let you know the next and final steps.

Lastly, if your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

You will receive an invoice from PLOS for your publication fee after your manuscript has reached the completed accept phase. If you receive an email requesting payment before acceptance or for any other service, this may be a phishing scheme. Learn how to identify phishing emails and protect your accounts at https://explore.plos.org/phishing.

If we can help with anything else, please email us at customercare@plos.org.

Thank you for submitting your work to PLOS One and supporting open access.

Kind regards,

PLOS One Editorial Office Staff

on behalf of

Dr. Dinesh Kumar Nishad

Academic Editor

PLOS One

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Supplementary Materials

    Attachment

    Submitted filename: Response to Reviewers.pdf

    pone.0358970.s002.pdf (103KB, pdf)
    Attachment

    Submitted filename: Response to Reviews Minor Revision.pdf

    pone.0358970.s003.pdf (71.2KB, pdf)

    Data Availability Statement

    All data and code required to reproduce the results in this study are publicly available in the GitHub repository at https://github.com/mohammadmgh78/Informed-Forecasting, with no access restrictions. The repository contains: (1) Jupyter notebooks implementing the full pipeline for each forecasting experiment described in the manuscript (influenza incidence, call-center arrival counts, and airline passenger counts), organized into data preparation, LLM-based forecasting, and result-analysis stages; (2) the raw and processed input time series used in each experiment; (3) cached outputs from all LLM forecasting queries, one file per experimental configuration, allowing the result-analysis notebooks to be rerun without repeating the API calls; and (4) the aggregated result tables and figures presented in the manuscript.


    Articles from PLOS One are provided here courtesy of PLOS

    RESOURCES