Table 2. Social media sentiment can lead financial returns. For each instrument above we show its largest statistically-significant information surplus seen in the study, i.e. Twitter sentiment's best ability to lead financial data ahead of time, relative to no time-shift. For each instrument, we also offer a summary of: the search characteristics of their Twitter Filters; their mean minutely message volume over the 3-month collection period; and their corresponding largest statistically-significant information surplus. We also demonstrate the leading time-shift (in hours) at which this occurs, and the corresponding sentiment type (positive, negative or net). We also report the total number of statistically-significant instances where social media sentiment leads financial data. Note: as discussed in the Methods, the full 24-hour autocorrelation-removal moving mean windows have been used throughout. We observe that Twitter Filter #11 (“$AAPL”) is the only filter admitted which uses just the financial instrument's industry Ticker-ID. *: We witness unexpectedly-low hourly volumes of string-unfiltered US Tweets. This is because we employed the most-accurate location-detection methodology available: only admitting those Tweets which are stamped with geographical-coordinates encompassed within the extremes of the United States' border. The majority of Tweets are not stamped with geographical-coordinates since typically only those messages which are sent from GPS-enabled devices may contain geographical-coordinates. Nonetheless, this hourly message volume was sufficient to pass our minimum mean message volume threshold of 1 message per minute. Finally, we note that our methodology identifies the following financial-instrument/Twitter-Filter combinations as inadmissible due to a lack of statistical-significance: Microsoft CFDs, FTSE100 CFDs and Futures, S&P500 CFDs, IBM CFDs, Wal-Mart CFDs and Bank of America CFDs. These assets do attract sufficient Tweet volumes, but their sentiments are not able to lead financial data in a statistically-significant manner for any of the leading time-shifts considered in this investigation.
| # | Instrument name | Twitter Filter | Mean message volume per minute | Largest statistically-significant information surplus |
|---|---|---|---|---|
| 1 | Apple, Inc. CFDs | $AAPL AND/OR “Apple” | 126.7 | 0.140% |
| 2 | Amazon.com, Inc. CFDs | $AMZN AND/OR “Amazon” | 123.1 | 3.473% |
| 3 | Google, Inc. CFDs | $GOOG AND/OR “Google” | 184.0 | 2.638% |
| 4 | Intel, Inc. CFDs | $INTL AND/OR “Intel” | 12.9 | 1.414% |
| 5 | Coca-Cola, Co. CFDs | $KO AND/OR “Coca Cola” AND/OR “Coca-Cola” | 24.8 | 0.723% |
| 6 | McDonald's, Corp. CFDs | $MCD AND/OR “McDonald's” AND/OR “McDonalds” | 46.5 | 1.902% |
| 7 | S&P500 Futures | String-unfiltered US Tweets | 142.7* | 2.462% |
| 8 | Oracle, Corp. CFDs | $ORCL AND/OR “Oracle” | 5.0 | 0.363% |
| 9 | Cisco Systems, Inc. CFDs | $CSCO AND/OR “Cisco” | 4.0 | 2.766% |
| 10 | The Home Depot, Inc. CFDs | $HD AND/OR “Home Depot” | 1.9 | 2.813% |
| 11 | Apple, Inc. (Ticker only) CFDs | $AAPL | 1.8 | 3.347% |
| 12 | J.P. Morgan, Inc. CFDs | $JPM OR “JPMorgan” OR “JP Morgan” | 1.1 | 3.936% |
| # | Instrument name | Leading time-shift corresponding to the largest information surplus | Sentiment type corresponding to the largest information surplus | Number of statistically-significant leading information surplus time-shifts |
| 1 | Apple, Inc. CFDs | 10 | Negative | 2 |
| 2 | Amazon.com, Inc. CFDs | 20 | Net | 30 |
| 3 | Google, Inc. CFDs | 14 | Net | 14 |
| 4 | Intel, Inc. CFDs | 1 | Negative | 2 |
| 5 | Coca-Cola, Co. CFDs | 8 | Positive | 13 |
| 6 | McDonald's, Corp. CFDs | 13 | Net | 7 |
| 7 | S&P500 Futures | 22 | Net | 1 |
| 8 | Oracle, Corp. CFDs | 1 | Net | 1 |
| 9 | Cisco Systems, Inc. CFDs | 13 | Net | 15 |
| 10 | The Home Depot, Inc. CFDs | 11 | Positive | 8 |
| 11 | Apple, Inc. (Ticker only) CFDs | 14 | Negative | 2 |
| 12 | J.P. Morgan, Inc. CFDs | 12 | Positive | 2 |