Inference limits in partially observable Ethereum blockchains

Abstract

While full ledger access is theoretically possible on public blockchains, in reality it is often not possible. Things that can be seen are limited by storage limitations, client design, indexing services, and off-chain execution pathways. This means that entire ledger objects are rarely used for empirical blockchain analysis; instead, observable projections are typically used. In this research, the observability of blockchain is recast as an inferential problem with incomplete observation. Studying identifiability, information loss, and irreducible uncertainty under coarsened access, the framework defines a full ledger, an observable ledger, and an observability mechanism. Three distinct visibility regimes, independent Bernoulli, clustered, and activity-dependent, are assessed in the simulation study. Reduced visibility raises uncertainty inflation, root mean squared error, variance, and mean squared error across all three regimes. The most severe deterioration happens when the condition of the underlying ledger determines visibility. This empirical study employs Google BigQuery’s publicly indexed Ethereum block data spanning blocks 18,000,000 to 18,001,000. Over the chosen Ethereum period, descriptive summaries reveal a large amount of fluctuation in gas utilised, transaction count, and basic charge per gas at the block level. Experiments with controlled missingness on the observed slice reveal that RMSE and trend estimate bias grow with increasing missingness, and that the degree of distortion is significantly affected by whether the incompleteness is MCAR-like, MAR-like, or MNAR-like. This research proves that partial observability isn’t just a secondary data issue; it can significantly affect inference on Ethereum block-level summaries.

Introduction

Public blockchains are decentralised databases that record transactions on ledgers that are meant to be visible to everyone and are designed to be additive only. Actually, who may view the whole ledger is dictated by system limits, not just protocol requirements. Storage constraints, client architecture, indexing services, and off-chain execution pathways determine the visible parts of the ledger at any given moment in time. Because of this, empirical blockchain analyses do not employ exact replicas but rather incomplete representations of the on-chain state. Since observed data often serve as a representation of the whole system, this differentiation is vital for auditing, detecting anomalies, evaluating performance, and reviewing policies.

All information may be visible, according to early blockchain models. The ledger was thought to be both globally accessible and locally verifiable according to the fundamental protocols, which meant that all participants had to keep detailed records of all transactions and state changes^1,2. Despite having existed before networks reached their present size, this paradigm made formal consideration of security and consistency easier. With the growth of blockchains came higher costs for storage and connectivity. Many nodes stopped keeping full ledgers when participation shifted to include partial replication and lightweight clients. This change and the subsequent restrictions on ledger visibility for nodes and outside observers have been documented in empirical and architectural studies^3–5.

The statistical ramifications of less exposure are huge. In controlled trials, the estimator variance increases from about 0.0050 to 0.0064, uncertainty inflation rises to 1.28, and bias is modest when observable coverage is reduced from full visibility (π = 1.0) to 80% (π = 0.8). With 60% visibility (π = 0.6), the uncertainty inflation is 1.66, and the variance is more than 0.0083. Uncertainty inflation reaches 2.5 times full observability and variation more than doubles when visibility drops to 40%. On average, less than 41 components are visible in the lowest visibility regime (π = 0.2). There is an increase in the uncertainty inflation to 5.0, the RMSE to0.14, and the variance to about 0.0253. Neither random noise nor the choice of an estimator can explain these patterns; rather, they result from data loss.

Users of blockchains often have limited visibility. Instead of directly accessing data, lightweight and superlight clients rely on indirect evidence from distant servers or concise proofs to ensure consistency via limited query interfaces that allow them to communicate with the ledger^6,7. As time goes on, ledger visibility fragmentation between participants increases due to pruning and state-minimization techniques that let nodes to function while discarding prior data^8,9.

This constraint is made clear by layer-2 rollup designs. Only compressed summaries or proofs are included in the base layer, and execution is moved off-chain. While this approach makes scaling easier, it makes direct auditing more complicated. Since observers in these systems rely on imperfect on-chain representations to form conclusions, data availability is crucial for inference¹⁰. Even still, a number of empirical investigations persist in treating observed ledger data as almost full, often supposing that missing entries only contribute to noise rather than structural uncertainty.

This setup is in line with missing-data theory, which states that the mechanism controlling the unseen components is what determines inference¹¹. Reconstruction could be badly organised, and more samples or more precise estimators won’t help when observation is limited¹². It has been shown in robust statistics and decision theory^13–15 that using coarser data raises risk irreducibly and cannot be mitigated by post hoc adjustment.

A common starting point for works on blockchain auditability is the idea that, since ledger data is publicly available, it should be possible to conduct direct empirical research with it. Storage constraints, client design, indexing services, and off-chain execution pathways all restrict what can be monitored, weakening this premise in practice. As a result of treating partial views as though they were virtually complete or as if missingness just added noise, statistical treatments of blockchain data fail to adequately address the gap between the hypothetical whole ledger and the accessible observable projection. By recasting blockchain observability as an inferential problem with incomplete observation, this paper fills that knowledge gap. An observable ledger, an observability mechanism, and a full ledger object form the basis of the analysis’s formal framework, which is then used to investigate issues related to identifiability, information loss, irreducible uncertainty, information visibility, and visibility-dependent variance inflation. Combining publicly indexed Ethereum block data from Google BigQuery with multi-regime simulation becomes the empirical strategy. Reduced visibility increases variance, mean squared error, root mean squared error, and uncertainty inflation; the simulation contrasts Bernoulli, clustered, and activity-dependent visibility regimes, with activity-dependent masking exhibiting the worst degradation. After that, the Ethereum study separates the descriptive variation in gas usage, transaction count, and base fee per gas from controlled-missingness validation on the observed slice. This demonstrates that trend-estimation bias and RMSE increase with missingness and are highly dependent on whether the incompleteness is MCAR-like, MAR-like, or MNAR-like. The paper addresses a methodological gap in the scientific literature by demonstrating, theoretically and empirically, how limited availability changes the reliability of blockchain-based statistical conclusions and by elevating blockchain observability from an implicit background assumption to an explicit inferential object.

Methodology

Let denote the full block-level ledger object over the selected Ethereum interval, with variables including gas used, transaction count, and base fee per gas. Let denote the observable ledger returned by the selected public access pathway. The formal inferential target is defined on , but the empirical analysis is performed on . The distinction between these two objects formalizes the observability constraint studied throughout the paper.

The notation and its correspondence to the simulation and Ethereum empirical analysis are summarized in Appendix Table A.

Spaces, objects, and observation operator

Let be a probability space. Let denote the space of full ledger realizations and let denote the space of observable ledger realizations. In the present study, the ledger is defined at block level over a fixed Ethereum block interval. For each block index t∈ I, let

Where is gas used, is transaction count, and is base fee per gas. The full ledger is the random element

1

Let be a class of observability mechanisms. Each acts as a measurable map

2

Thus, M maps the full ledger object to the ledger object available for inference. In the empirical analysis, represents the block-level data returned by the selected public access pathway. The framework does not assume that all ledger components outside this pathway are observed.

To allow observer-specific or time-varying access conditions, M may itself be random. In that case,

The corresponding observable information is

Definition 1

(Partially observable blockchain). The ledger is partially observable under M if

3

Remark 1

(Full observability). If , where for all , then

Visibility mask and missingness structure

Let denote an index set of ledger components. In the general framework, these components may represent blocks, block-level fields, transactions, receipts, logs, traces, witnesses, or other ledger objects, depending on the access model under study. For the empirical analysis in this paper, the observed object is restricted to block-level variables over the selected Ethereum block interval¹⁷.

Define the random visibility mask

where

For a realized ledger , define the induced observable ledger by

4

Thus, the observable ledger is the restriction of the full ledger to the visible index set, and

Definition 2

(Observability mechanism). An observability mechanism is a random mask together with the induced projection . Its stochastic behavior is characterized by the conditional visibility probabilities

5

or, more generally, by the conditional law.

The collection describes how access constraints map the full ledger to its observable projection.

Remark 2

(Connection to missing-data mechanisms). The visibility process is of MCAR type if

6

and of MAR type if

7

These are formal benchmark cases. They are introduced to classify observability regimes and not to claim that empirical blockchain access constraints necessarily satisfy MCAR or MAR.

Targets, estimators, and reliability functionals

Let . The inferential target is a functional of the full ledger,

An estimator is any -measurable map of the observable ledger,

For a fixed observability mechanism M, define the estimation error by

The corresponding bias, covariance, and mean squared error are

and

A scalar reliability functional may then be written as

or, more generally, as a weighted combination of squared bias and total variance,

To compare uncertainty under partial observability with the full-visibility benchmark , define the uncertainty inflation index

8

Definition 3

(Audit functional). Let

be a scalar functional defined on the full ledger. In the present framework, A represents any scalar summary whose stability under reduced observability is of interest. This includes, for example, block-level rolling summaries derived from gas used, transaction count, or base fee per gas.

Define

The audit deviation induced by mechanism M is

When a bounded reporting scale is useful, define

The quantity is the theoretical object of interest. The transformation AC(M) is retained only as an optional reporting index.

Remark 3

The audit functional is introduced to connect the abstract framework to empirical summaries. In the empirical section, the relevant summaries are rolling dispersion and related block-level indicators computed from the observed Ethereum series. No claim is made that AC(M) is a primary inferential quantity.

Identifiability (partial visibility)

Definition 4

(Identifiability under partial observability). Let be an observability mechanism. For each realized ledger , define the observational equivalence class

A parameter is identifiable under M if

Thus, identifiability holds when the observable projection determines the target uniquely up to null sets.

Proposition 1

(Information loss monotonicity). Let . If is coarser than , in the sense that

9

then for any square-integrable estimand ,

10

Proof

Under squared loss, the optimal estimator among -measurable functions is and the corresponding minimum risk equals ¹⁵. Since , conditional variance is monotone under refinement, which yields Eq. (10). Equivalently, Blackwell-dominates , so the coarser experiment cannot achieve uniformly smaller squared-error risk¹⁴.

Proposition 2

(Conditional-variance lower bound). Let with Then

11

Proof

By the projection identity,

The first term is nonnegative and vanishes when . Hence the infimum equals the conditional-variance term^12,15.

Definition 5

(Irreducible uncertainty). The irreducible uncertainty induced by mechanism M is

12

This is the minimum achievable mean-squared error under squared loss among all estimators measurable with respect to the observed ledger.

Scope constraint

The analysis is conducted entirely on the observable ledger . No reconstruction of the full ledger from is attempted, and no surrogate object is introduced as a recovered version of the hidden ledger.

The purpose of the framework is therefore not to recover unobserved ledger components. It is to characterize, for a given observability mechanism M, the inferential quantities induced by partial visibility, including

The framework treats observability as part of the inferential setting and studies the statistical consequences of restricted access on estimation and uncertainty.

Visibility regimes and variance inflation

We first consider an independent Bernoulli visibility regime as a benchmark case. This regime is introduced for analytical tractability and baseline comparison. It does not represent the full range of structured observability constraints that may arise in blockchain data systems.

Let be a component-wise visibility mask with independent entries

Let denote the visibility profile. The induced observability mechanism is

Proposition 3

(Variance inflation under Bernoulli visibility). Let with . Under squared-error loss, the optimal estimator based on the observed ledger is The corresponding irreducible uncertainty is If , then

13

Proof

By Definition 5, irreducible uncertainty equals the expected conditional variance under the observed σ-field. If component wise, then yields a finer observable projection than . The corresponding observable σ-field is therefore richer, and conditional variance decreases under refinement. This gives Eq. (13).

Corollary 1

(Uniform visibility). If ∀ , define Then is non-increasing in π, with limits

14

Definition 6

(Variance inflation factor). Let denote full observability. Define the variance inflation factor by

15

Then

and, in the uniform-visibility regime, is non-increasing as π increases. A small-bias regime under independent Bernoulli visibility should therefore be interpreted as a property of this benchmark observability model. It does not imply that bias remains negligible under structured, clustered, or activity-dependent visibility constraints.

Figure 1 summarises the process by which access constraints drive the non-invertible projection from the whole blockchain ledger to an observable ledger state. Impairment in identification, coarsening of data, and missing components are the results of restricted access, which in turn causes inflation of variance, bias based on mechanisms, and irreducible uncertainty in inferences. These consequences are a direct result of the observability mechanism. Do not attribute them solely to stochastic noise or adversarial behaviour; estimator choice or more processing will not eradicate them.

Fig. 1.

Chart for conceptual framework of blockchain observability as a statistical inference problem.

Simulation design

Let be i.i.d. ledger realizations drawn from a common distribution ,

For each realization, define the full-ledger target

For a given observability regime, the observable ledger is generated by applying the corresponding mechanism M to each realization:

The associated estimator is

The simulation study evaluates three classes of observability regimes.

1. Independent Bernoulli visibility: Each component is retained independently with probability .

2. Clustered visibility loss: Consecutive components are removed in blocks, to represent bursty access failures or indexer interruptions.

3. Activity-dependent visibility: Visibility probabilities depend on the realized ledger state, so that highly active components may be more likely to be hidden than low-activity components.

The first regime is a baseline benchmark. The second and third regimes are included to reflect the possibility that blockchain observability constraints are structured rather than independent.

For each regime and each visibility level, the empirical mean estimator is

The empirical bias, variance, and mean squared error are defined by

and

Definition 7

(Empirical uncertainty inflation). Let denote full observability. The empirical uncertainty inflation under mechanism M is

16

In the Bernoulli benchmark regime, the visibility grid is taken as

17

For each regime and each visibility setting, the simulation reports

Under i.i.d. sampling and finite second moments, is a consistent empirical analogue of the corresponding variance-inflation quantity. In the independent Bernoulli benchmark, it is expected to increase as visibility decreases. Under clustered or activity-dependent visibility, the same monotone pattern need not preserve the same bias profile, and this distinction is part of the simulation design rather than an exception to it.

Simulation

In this simulation study, we appear at how changing the observation mechanism alone affects inference while the full-ledger target stays the same. We computed the target functional from the full ledger after generating a ledger realisation with (m = 200) components for each replication. Afterwards, three visibility regimes, activity-dependent visibility, clustered visibility loss, and independent Bernoulli visibility, were used to enforce partial observability. The standard example is the Bernoulli regime. The clustered regime stands for omission that is either bursty or block-correlated. Highly active realisations may become less observable than low-activity realisations due to the activity-dependent regime, which reflects observability loss that depends on the realised ledger state. In “Methodology” section, the methodological framework is matched by this configuration, where observability is considered an integral component of the inferential environment instead of an after-the-fact disruption.

The visibility levels π = (1.0, 0.8, 0.6, 0.4, 0.2) were assessed in the simulation, which utilised (N = 3000) realisations. Uncertainty inflation, variance, empirical bias, and root mean squared error were calculated for each visibility level and regime. Figure 2 summarises the combined behaviour of variance, mean squared error, root mean squared error, and uncertainty inflation, whereas Table 1 reports the complete numerical data. Separate from each other, Fig. 3 displays the corresponding bias patterns.

Fig. 2.

Simulation metrics under alternative visibility regimes plot.

Table 1.

Simulation metrics across visibility levels under Bernoulli, clustered, and activity-dependent observability regimes for m = 200.

Regime	π		AOC	Bias	Var	MSE	RMSE	Var full	UI
Activity-dependent	1	3000	196.0387	0.0003	0.0051	0.0001	0.0101	0.0051	1.0000
	0.8	3000	149.1207	0.0018	0.0081	0.0030	0.0552	0.0051	1.5791
	0.6	3000	115.6690	− 0.0011	0.0137	0.0089	0.0943	0.0051	2.6641
	0.4	2999	84.5869	− 0.0017	0.0260	0.0212	0.1456	0.0051	5.0716
	0.2	2997	50.9296	− 0.0020	0.0537	0.0490	0.2215	0.0051	10.4646
Bernoulli	1	3000	200.0000	0.0001	0.0050	0.0001	0.0001	0.0050	1.0000
	0.8	3000	160.1440	− 0.0001	0.0064	0.0013	0.0355	0.0050	1.2745
	0.6	3000	120.0487	0.0006	0.0088	0.0034	0.0583	0.0050	1.7562
	0.4	3000	80.1077	− 0.0017	0.0128	0.0078	0.0881	0.0050	2.5520
	0.2	3000	39.9893	− 0.0015	0.0248	0.0202	0.1422	0.0050	4.9647
Clustered	1	3000	200.0000	0.0001	0.0050	0.0001	0.0001	0.0050	1.0000
	0.8	3000	157.9420	0.0007	0.0063	0.0014	0.0369	0.0050	1.2670
	0.6	3000	116.4060	− 0.0002	0.0086	0.0036	0.0602	0.0050	1.7273
	0.4	3000	76.7513	− 0.0003	0.0132	0.0081	0.0902	0.0050	2.6413
	0.2	3000	37.4150	− 0.0002	0.0268	0.0217	0.1474	0.0050	5.3692

Fig. 3.

Bias across visibility regimes plot.

Degradation trends are comparable in the Bernoulli and clustering regimes. Table 1 shows that both begin with an uncertainty inflation of 1,000,000 and a variance of close to 0.005 when both are fully visible. Figure 2's four panels show that in both regimes, variance, mean squared error, root mean squared error, and uncertainty inflation all rise monotonically as visibility decreases. Table 1 shows that variance increases from 0.0050 at π = 1 to 0.0248 at π = 0.2 in the Bernoulli regime, while uncertainty inflation grows from 1.0000 to 4.9647. In the clustered regime, variance rises from 0.0050 to 0.0268 and uncertainty inflation from 1.0000 to 5.3692, again as reported in Table 1 and reflected in the corresponding trajectories in Fig. 2. These findings demonstrate that the independent benchmark is not the only place where the inflation pattern is observed. It remains even when the loss of vision is organised over different parts.

If visibility is low, the activity-dependent regime will cause inferential degradation to be much stronger. The uncertainty inflation goes up from 1.0000 to 10.4646, and the variance goes up from 0.0051 at π = 1 to 0.0537 at π = 0.2, as shown in Table 1. The behaviour of the root mean squared error and the mean squared error are identical. From 0.0101 at full visibility to 0.2215 at π = 0.2, the root mean squared error increases, as illustrated in Table 1. The activity-dependent curve stays above the Bernoulli and clustered curves across the visibility grid in all four panels of Fig. 2, indicating this stronger degradation. The comparison shows that independent censoring is not always the best strategy, and that observability loss associated with the realised ledger state can lead to far higher uncertainty inflation.

While variations in variance and mean squared error have a larger impact on the extent of bias, the bias profile varies depending on the regime. At π = 0.2 and 0.0006 at π = 0.6 in Table 1, the bias varies close to zero in the Bernoulli regime across the visibility grid. As can be seen in Table 1 and Fig. 3, the deviations are similarly modest and practically nonexistent in the clustered regime. According to Table 1 and Fig. 3, the absolute values of bias range from (− 0.0020) at π = 0.2 to 0.0018 at π = 0.8 in the activity-dependent regime, which is bigger than in the benchmark setting. The sign of the bias changes as well. It is not possible to establish as a mechanism-free conclusion the preceding assertion that bias is still limited; this is especially true for the Bernoulli benchmark.

Combined, the findings from the simulation lend credence to two arguments. For starters, there isn’t just one idealised masking scheme that might cause inferential instability under reduced visibility; all three regimes in Fig. 2 and Table 1 exhibit the same monotonic worsening of variance, mean squared error, root mean squared error, and uncertainty inflation. Secondly, the observability mechanism’s structure determines the magnitude of the instability. Figure 2 and Table 1 show that, at low coverage, activity-dependent visibility causes the greatest inflation in uncertainty and error, while clustered visibility causes the mildest deterioration and a closely related but slightly stronger pattern. Further evidence that low bias is not universal across visibility techniques is shown by the bias behaviour in Fig. 3. Results like these support the methodological argument that observability isn’t a secondary data limitation but rather an inferential difficulty that needs modelling.

Ethereum data analysis

For blocks 18,000,000 to 18,001,00, we analysed Ethereum block data collected using Google BigQuery. In this part, the concrete item is the observable series at the block level that the chosen public access route returns. The gas used, transaction count, and base charge per gas are block-level variables that make up this observable object in the present investigation. It omits features like receipts, logs, traces, archival state, and delayed-access recovery of hidden components, as well as transaction-level reconstruction. We skip reconstructing the whole ledger and instead do the analysis on the observed projection. Differences in petrol use, number of transactions, and base fee per petrol are summarised in the descriptive component. The validation part measures the distortion of inference caused by partial observability by applying controlled incompleteness to the observed slice.

The three series’ observed variability at the block level is summarised in Fig. 4. In the timeframe under consideration, the observed block slice does not exhibit variance stability, as demonstrated by the rolling-dispersion panels. In contrast to discrete variations, the raw-series panels reveal continuous shifts in level and spread. Descriptive plots are what these are. They prove that Ethereum’s observable series vary across the chosen block range, but they can’t tell you for sure if the instability is due to on-chain intrinsic temporal variation or partial observability.

Fig. 4.

Observable Ethereum block-level variability across gas used, transaction count, and base fee per gas.

We used three different methods, MCAR, MAR-like, and MNAR-like, to apply controlled missingness to the observed Ethereum slice in order to directly evaluate inferential distortion. We compared the estimated slopes of the linear trends for gas utilised, transaction count, and base charge per gas with the full-view slope obtained from the observed slice without extra masking. We repeated this process for each mechanism and missingness rate. The results of the (n = 250) Monte Carlo replications for each setting are reported in Table 2, together with the RMSE and bias values. The RMSE profiles are summarised in Fig. 5, while the bias profiles are shown in Fig. 6.

Table 2.

Controlled-missingness distortion in ethereum block data for n = 250.

Metric	Mechanism	Miss rate (%)	Bias	RMSE
Base fee per gas	MAR-like	10	− 8618	102,900
		20	− 28,570	159,400
		30	− 38,750	215,100
		5	− 10,420	68,430
		50	− 146,100	354,400
	MCAR	10	3710	96,220
		20	15,050	151,000
		30	4508	177,800
		5	3832	64,840
		50	20,090	281,700
	MNAR-like	10	176,900	222,900
		20	476,900	519,300
		30	962,300	1,015,000
		5	75,460	125,300
		50	3,290,000	3,303,000
Gas used	MAR-like	10	7.242	192
		20	34.29	268.1
		30	128.3	461.7
		5	8.261	134.1
		50	383.5	663
	MCAR	10	− 6.561	184.2
		20	15.94	260.2
		30	− 14.12	323.3
		5	− 1.644	125.7
		50	− 44.29	469
	MNAR-like	10	− 73.85	243.1
		20	− 145.1	361.5
		30	− 337.5	550.4
		5	− 16.6	150.1
		50	− 478.5	595.9
Transaction count	MAR-like	10	0.0002069	0.001526
		20	0.0005136	0.002536
		30	0.001112	0.003282
		5	5.55E-05	0.00117
		50	0.001774	0.003765
	MCAR	10	− 7.52E-05	0.001394
		20	3.48E-05	0.002083
		30	4.65E-05	0.002592
		5	− 0.0001143	0.0009164
		50	4.82E-05	0.004258
	MNAR-like	10	− 0.0001462	0.00193
		20	− 6.86E-05	0.002611
		30	0.0002146	0.003316
		5	− 0.0001351	0.001162
		50	0.002122	0.003519

Fig. 5.

RMSE of linear trend estimates under controlled Missingness.

Fig. 6.

Bias of linear trend estimates under controlled Missingness.

With the exception of MNAR-like masking, the RMSE results for all three variables rise as missingness increases. As seen in Table 2 and Fig. 5, the root-mean-squared error (RMSE) for the basic fee per petrol increases from 64,840 at 5% MCAR missingness to 281,700 at 50% MCAR missingness, with a corresponding MNAR-like increase from 125,300 to 3,303,000. The pattern for gas utilised is identical. Under MNAR-like masking, the RMSE increases from 150.1 to 595.9, as shown in Table 2 and traced in Fig. 5, and it goes up from 125.7 at 5% MCAR to 469.0 at 50% MCAR. The monotonic degradation is also apparent for transaction count, but on a smaller numerical scale. Figure 5 shows that the mechanism-specific differences persist even after adjusting the RMSE from 0.0009164 at 5% MCAR to 0.004258 at 50% MCAR. The results demonstrate that the distortion caused by decreased visibility is affected by both the extent of missingness and the mechanism by which it is generated.

Additionally, the bias results vary depending on the method. Table 2 and Fig. 6 show that, under MCAR, the bias for all three variables stays near zero for the majority of missingness rates. The bias becomes more apparent with MAR-like masking, especially for gas utilised and base fee per gas calculations. The bias increases significantly when MNAR-like masking is applied. The basic fee per petrol shows an MNAR-like bias of 3,290,000 at 50% missingness compared to 75,460 at 5% missingness (Table 2 and Fig. 6). Table 2 and Fig. 6 show that the MNAR-like bias changes for gas utilised from -16.6 at 5% missingness to − 478.5 at 50%. Although there is less absolute bias in the transaction count, the same mechanism reliance is still there. So, it’s not entirely mechanism-free to say that bias remains modest. Under MAR-like and MNAR-like incompleteness, it becomes less stable, but it is most compatible with the MCAR benchmark.

The findings of the validation and descriptive analyses, taken together, lend credence to a more limited empirical conclusion. As seen in Fig. 4, the amounts accessible through the public access method show a great deal of diversity in the reported Ethereum block slice. As demonstrated in Table 2, Figs. 5, and 6, the bias and RMSE both rise when controlled incompleteness is applied to that observed slice. The magnitude of this rise is highly dependent on whether the masking mechanism is MCAR-like, MAR-like, or MNAR-like. Partial observability can significantly affect inference on Ethereum block-level summaries, according to the real-data evidence. The more robust assertion that observability limitations are the sole source of volatility in the raw series is unnecessary.

Discussion and conclusions

In real-world blockchain analysis, full ledger visibility is unusual. Nothing can be accessed due to storage limitations, client architecture, indexing layers, or off-chain execution paths. So, rather than being a complete record, the observable ledger is a projection. This projection, and not just the estimator used to it, shapes inference. Data that is not visible to the naked eye cannot be recovered by further computation or local refining.

Why this matters theoretically

According to the findings, inferential dependability revolves around the observation process. It becomes more difficult to identify an individual and the minimum acceptable risk rises when multiple complete ledgers point to the same observable state. In such a case, the observation structure introduces uncertainty. This impact is independent of out-of-the-ordinary actions or selection of estimators. The available information is coarsened, and this is the result.

Results from computer models

There is a consistent pattern of degradation in low visibility, according to the simulation data. As visibility reduces, uncertainty inflation, root mean squared error, variance, and mean squared error all rise throughout clustered, Bernoulli, and activity-dependent regimes. The masking structure determines the magnitude of the same directional trend that occurs across regimes. The most severe degradation is observed in the activity-dependent regime. In the clustered regime and the Bernoulli benchmark, bias is quite minimal; however, under activity-dependent visibility, it becomes more mechanism-dependent. While demonstrating that bias behaviour is not mechanism-free, the simulation evidence also lends credence to a structural interpretation of inferential deterioration.

Proof from Ethereum public ledger information

Within a narrower empirical context, the Ethereum study lends credence to the identical result. Variation in gas utilised, transaction count, and base cost per gas is evident in the observed block-level slice. These patterns of description show that the observable series are different over the chosen interval, but they don’t tell us anything about where the variation is coming from. The more robust empirical outcome is provided by the controlled-missingness approach. The bias and root-mean-squared error (RMSE) rise when the observed Ethereum slice is incomplete; the magnitude of this rise is highly dependent on the type of masking method used, which can be MCAR, MAR, or MNAR. For base charge per gas and gas used, the biggest distortion happens under MNAR-like masking. Therefore, the real-data evidence demonstrates that inference on Ethereum block-level summaries can be significantly impacted by partial observability.

How blockchain analytics are affected

Because imperfect access is sometimes handled too carelessly, these findings significance for empirical blockchain research. It is common practice to treat observed ledger data as near entire records when conducting analyses, even when access is mediated through indexers, lightweight clients, or partial protocol views. Both the seeming stability and the level of uncertainty might be inflated by this. Accordingly, the access mechanism that produced the data should be considered alongside auditability claims when interpreting the data. When complete replication is not possible, a more solid foundation for empirical research can be established by treating blockchain observability as an inferential constraint.

The study concludes that blockchain analytics do not suffer from secondary annoyances caused by partial observability. This limitation is fundamental to inferential logic. Information loss reduces identifiability and increases uncertainty; the theoretical results show how this happens. This effect is observed in both the benchmark and structured observability regimes, according to the simulation and Ethereum analyses. Understanding the statistical findings supported by blockchain technology requires keeping the observable projection in mind.

Limitations and future work

Intentionally, the study’s scope is small. Absent any effort to recover concealed ledger states, the study focuses on the inferential consequences of limited observability. While this does make it easier to spot the statistical impact of coarsened access, it does so at the expense of considering all aspects of blockchain systems that could influence the data that is actually observed.

Capacity for modelling

Data is collected and analysed at the block level. It doesn’t treat receipts, logs, execution traces, archival state, mempool dynamics, or ordering at the transaction level as independent inferential objects. Therefore, the current design does not account for intra-block dependencies or the behaviour of miners during execution. Not only are the types of information loss discussed above not accounted for, but there may be additional types due to data propagation effects and delays in communication across layers.

Evidence-based coverage

The Ethereum study can only cover a small range of consecutive blocks. This window describes long-run network behaviour or greater structural changes across time, yet it is enough to show block-level variability and validate inferential distortion under controlled incompleteness. Hence, the empirical results are not meant to represent a comprehensive explanation of Ethereum dynamics, but rather proof of inferential sensitivity under partial observability during the chosen interval part.

Future work

There are a number of open extensions. One approach is to be more explicit when modelling observability mechanisms that are time-varying, field-correlated, or client-specific. Second, we may build on top of the current architecture to support more advanced blockchain objects like archival state views, traces, logs, and receipts. There is an operable third direction. According to Ressi et al.¹⁶, AI-assisted methods have the potential to aid in observing profiles, detecting indexing drift, supporting observability-aware analytics, and calibrating uncertainty based on access quality. However, it is important to note that these methods do not guarantee complete reconstruction of concealed ledger states. The link between statistical inference and real-world blockchain data systems would be strengthened by these extensions.

Implementation

The observable ledger is the sole platform on which the framework is applied. In order to retrieve previously hidden data, no surrogate ledger is used, and no whole ledger states are restored. Not a modelling convenience, but the access pathway defines the difference between the complete ledger and the observable ledger. For every calculation, we only use data that is truly accessible via the chosen interface.

The verifiable ledger projection is the basis for the analytical process. Here, we use this forecast as the returned data at the block level from Ethereum’s chosen public access channel. Outside of that perspective, it is not presumed that full ledger states are available. Therefore, the data set used for inference is decided by the access model.

The observed block-level series are the only outcomes of any statistical techniques. We do not reconstruct, impute, or backfill any missing blocks, transactions, or state variables. At the validation stage, we see how inference changes when visibility degrades under MCAR, MAR-like, and MNAR-like methods by applying controlled incompleteness to the observed Ethereum slice. This maintains coherence between the inferential model and the data that was actually analysed.

This same implementation idea can be used in several blockchain environments where access is limited due to client design, storage limits, indexing technique, or protocol architecture. Modifying protocols or replicating archives in their whole is not necessary for the framework. In the context of actual operational limitations, it is responsible for quantifying the impact of restricted access on inferential dependability.

Appendix

See Table 3.

Table 3.

Notation and empirical mapping.

Symbol/concept	Formal meaning	Simulation/Ethereum instantiation
	Underlying probability space	Probability space supporting ledger realizations and observability mechanisms
	Space of full blockchain ledgers	Full simulated ledger space; block-level Ethereum ledger object over the selected interval
	Space of observable ledgers	Masked simulated ledger space; observable BigQuery block-level projection
	Random full ledger	Full simulated realization; block-level Ethereum slice treated as the full empirical object for the present study
	Block at time index (t)	Ethereum block in the selected interval
	Gas used in block (t)	BigQuery gas-used field
	Transaction count in block (t)	BigQuery transaction-count field
	Base fee per gas in block (t)	BigQuery base-fee field for post-EIP-1559 blocks
	Class of observability mechanisms	Bernoulli, clustered, and activity-dependent simulation regimes; MCAR, MAR-like, and MNAR-like controlled masking in Ethereum validation
	Observation operator / ledger–observer mapping	Masking rule applied to the full ledger or observed Ethereum slice
	Observable ledger induced by (M)	Masked simulated ledger; observable Ethereum block-level series returned by the selected access pathway
	Block-level state at index t	Gas used, transaction count, and base fee per gas in block t
	σ-field generated by	Information available after masking or restricted access
	Index set of ledger components	Component positions in simulation; block-level variables or observations in the Ethereum slice
	Visibility mask	Binary mask over simulated components or over Ethereum observations in controlled-missingness validation
	Component (i) is observable	Retained simulated component or retained Ethereum observation
	Component (i) is hidden	Masked simulated component or removed Ethereum observation
	Visibility probability of component (i)	Component-wise retention probability
	Visibility profile	Visibility level in simulation, e.g. ({1.0,0.8,0.6,0.4,0.2})
	Estimand defined on the full ledger	Target functional in simulation; linear trend slope in Ethereum validation
	Estimator based on	Estimator computed after masking or restricted observation
	Irreducible uncertainty under mechanism (M)	Minimum achievable squared-error risk under the selected visibility mechanism
	Uncertainty inflation relative to full observability	Variance ratio in simulation
	Variance inflation factor	Visibility-dependent variance inflation under Bernoulli benchmark
	Identity / full-visibility mechanism	No masking; uncensored benchmark
	Audit or summary functional	Rolling dispersion or other block-level summary of gas used, transaction count, or base fee per gas
	Audit deviation under (M)	Difference between full and observed summary under the selected mechanism
MCAR	Missing completely at random	Controlled Ethereum masking independent of the target variable
MAR-like	Missingness depending on observed covariates	Controlled Ethereum masking depending on observed block-level quantities
MNAR-like	Missingness depending on the target itself	Controlled Ethereum masking depending on the variable being inferred

Author contributions

M.Z.A. conceived the study, developed the theoretical framework, carried out the statistical analysis, and wrote the manuscript. A.A. contributed to the protocol-level perspective, interpretation of the results, and critical revision of the manuscript. Both authors reviewed and approved the final manuscript.

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Data availability

The analysis uses publicly indexed Ethereum blockchain data representing an observable projection of the on-chain ledger. Specifically, data were obtained from the Google BigQuery *Ethereum Cryptocurrency* dataset, accessed through the bigquery-public-data.crypto_ethereum tables. The empirical study focuses on blocks 18,000,000–18,001,000. All results can be reproduced using the SQL queries described in the Methods section.

Code availability

The custom R code and supporting algorithmic workflow used to generate the simulation results, Ethereum analyses, tables, and figures central to this study have been archived in Zenodo and are available at 10.5281/zenodo.20094031. The repository includes the real-data Ethereum analysis script, the multi-regime simulation script, documentation, and instructions required to reproduce the reported results.

Declarations

Competing interests

The authors declare no competing interests.