Do Prediction Market Whale Signals Precede Equity Sector Returns?

Abstract

This paper investigates whether large probability shifts ("whale signals") on the Polymarket prediction market platform precede abnormal returns in correlated equity sector exchange-traded funds (ETFs). Using 604 resolved markets, 2,623 detected signal events, and 16,731 signal–ETF observations spanning January 2023 to December 2024, I apply standard event study methodology to test whether prediction market order flow contains information not yet reflected in equity prices. The broad hypothesis — that whale signals predict positive rotation toward beneficiary sectors — is rejected: signals are associated with mild, statistically significant negative abnormal returns across nearly all sectors, consistent with sophisticated participants using prediction markets to hedge equity exposure rather than to express naive directional views. Within this null result, three robust sub-findings emerge. First, geopolitical signals show genuine directional predictability of 55.5% at a seven-day horizon (p = 0.0008) and a thirty-day cumulative average abnormal return (CAAR) of +1.43% (p = 0.0037) — the only positive sector CAAR in the dataset. Second, cryptocurrency signals exhibit strong reverse predictability: equities predicted to benefit (MSTR, COIN, MARA, RIOT) instead decline by a thirty-day CAAR of −16.46% (p < 0.0001, n = 1,028), implying that large bullish prediction-market positions signal informed equity selling. Third, medium-confidence sector predictions outperform high-confidence predictions (52.3% vs. 46.9% directional accuracy at seven days), consistent with obvious first-order effects being pre-priced while subtler second-order effects remain partially unexploited. A second study replacing category-level sector mapping with entity-level reasoning (specific policy positions and geographic supply chains) significantly degraded accuracy in every category (overall seven-day accuracy fell from 46.4% to 43.6%, p < 0.0001), suggesting that prediction markets transmit broad categorical uncertainty to equities rather than entity-specific information, which is already priced. All data derive from public APIs requiring no authentication, ensuring full reproducibility.

Keywords: prediction markets, alternative data, event study, abnormal returns, market efficiency, Polymarket

1. Introduction

The efficient market hypothesis holds that asset prices incorporate all available information (Fama, 1970). Yet a substantial literature documents that information from non-traditional, or "alternative," data sources reaches equity markets with measurable delays, creating temporary windows of predictability (Katona, Painter, Patatoukas & Zeng, 2018; Chen, De, Hu & Hwang, 2019). Prediction markets are a distinctive alternative data source: unlike opinion polls or social-media sentiment, they aggregate the views of participants who stake real capital on being correct, which in principle should make their prices unusually information-dense (Wolfers & Zitzewitz, 2004).

This paper asks a specific question that, to my knowledge, has not been examined in the published literature: when a large participant places an outsized position on Polymarket and moves an event's implied probability sharply, do the equity sectors logically connected to that event subsequently earn abnormal returns? For instance, when a participant bets heavily that a military conflict will occur, do defense and energy ETFs outperform in the days that follow? The hypothesis is not that prediction markets forecast events — they demonstrably do that well — but that there exists an information transmission lag between prediction markets and equity markets that could be detected and, in principle, traded.

I test this in two stages. The first study (Study 1) detects whale signals across two years of resolved Polymarket data, maps each to affected sector ETFs using a category-level classifier, and measures abnormal equity returns via event study methodology. The second study (Study 2) replaces the category-level mapping with a more granular entity-level reasoning layer — incorporating specific political figures' policy positions and country-specific supply-chain effects — and asks whether this added specificity improves predictive accuracy.

The headline result of Study 1 is that the broad hypothesis fails: whale signals precede mild, broadly negative abnormal returns rather than the positive sector rotation the hypothesis predicts. However, three specific findings survive rigorous statistical scrutiny: geopolitical signals are directionally predictive on a one-week lag; cryptocurrency signals are predictive in the reverse direction with a large effect size; and medium-confidence predictions outperform high-confidence ones. Study 2 produces a second surprise: entity-level reasoning makes predictions worse, not better, which I argue reveals something fundamental about what kind of information prediction markets actually transmit to equity markets.

The remainder of this paper proceeds as follows. Section 2 reviews the relevant literature. Section 3 describes the data. Section 4 sets out the event study methodology. Sections 5 and 6 present the results of Study 1 and Study 2 respectively. Section 7 discusses interpretation and trading implications, Section 8 covers limitations, and Section 9 concludes.

2. Literature Review

2.1 Prediction Markets as Information Aggregators

The theoretical foundation for prediction markets rests on Hayek's (1945) argument that prices in competitive markets efficiently aggregate dispersed private information. Wolfers and Zitzewitz (2004) provide extensive evidence that prediction market prices function as well-calibrated probability estimates that frequently outperform expert forecasts and polls. Berg, Nelson and Rietz (2008) show that the Iowa Electronic Markets predicted U.S. presidential election outcomes more accurately than polls in the large majority of cases over two decades. Hanson (2003) formalizes market scoring rules that incentivize truthful information revelation, implying that large positions taken against prevailing prices carry disproportionate information content — the theoretical basis for treating "whale" activity as a signal.

2.2 Alternative Data and Equity Returns

The use of non-traditional data to predict equity returns expanded rapidly through the 2010s. Katona et al. (2018) demonstrate that satellite imagery of retail parking lots predicts earnings surprises. Chen et al. (2019) find that opinions transmitted through social media predict stock returns and earnings. Da, Engelberg and Gao (2011) show that Google search intensity predicts returns over roughly two-week horizons, and Preis, Moat and Stanley (2013) link search behavior to market movements. The common thread is partial efficiency: markets absorb widely disseminated information quickly but incorporate specialized alternative data more slowly. Prediction market order flow is a natural but under-studied addition to this category.

2.3 Event Study Methodology

The event study, introduced by Fama, Fisher, Jensen and Roll (1969) and formalized by MacKinlay (1997), is the standard framework for measuring equity price responses to discrete information events. The approach estimates expected "normal" returns over an estimation window, then computes abnormal returns as the deviation of realized from expected returns during an event window. It has been applied to earnings announcements (Ball & Brown, 1968), corporate control events (Jensen & Ruback, 1983), and monetary policy communications (Gürkaynak, Sack & Swanson, 2005). Applying it to prediction market signals extends this tradition to a new class of information event.

2.4 Informed Trading and Price Discovery

Kyle (1985) models how informed traders optimally disguise their activity to avoid moving prices against themselves, and Easley and O'Hara (1987) show how trade size conveys information to market makers. These models are directly relevant to the interpretation advanced in this paper: a sophisticated participant holding an equity position may use a prediction market to express or hedge a view in a venue where their activity is publicly visible, producing a signal whose naive interpretation is misleading. The cryptocurrency reversal finding in Section 5.3 is best understood through this lens.

2.5 Contribution

No prior published work, to my knowledge, links large-position activity on a decentralized prediction market to cross-asset equity sector returns using event study methodology. This paper provides that analysis on a large two-year sample and, through its two-study design, distinguishes between categorical and entity-specific information transmission.

3. Data

3.1 Prediction Market Data

Prediction market data were obtained from Polymarket via two public APIs, neither requiring authentication: the Gamma API (`gamma-api.polymarket.com`) for market metadata and the CLOB API (`clob.polymarket.com`) for price history. I retrieved all resolved markets with total trading volume of at least \$100,000 and resolution dates between January 1, 2023 and December 31, 2024. An initial pull returned 10,100 raw markets, of which 619 met the volume and date criteria and 604 had recoverable price history.

Price history was retrieved at hourly fidelity. A technical note: the CLOB endpoint's `interval=max` parameter returns empty histories for resolved markets, an undocumented limitation. The working approach paginates backward from each market's resolution date in thirteen-day windows using explicit `startTs` and `endTs` parameters, which reliably returns complete hourly series.

3.2 Signal Detection

A whale signal is defined as a movement in the YES outcome token price exceeding seven percentage points within any rolling sixty-minute window. This threshold balances sensitivity against noise. Where multiple qualifying movements occurred within six hours on the same market, only the earliest was retained, preventing a single information event from being counted multiple times.

This procedure identified 3,609 raw signals, reduced to 2,623 after deduplication. Signals were near-evenly split between upward movements toward YES (1,318) and downward movements toward NO (1,305), indicating no directional bias in detection. The mean absolute probability shift was 19.2 percentage points (median 14.4), ranging from the 7-point floor to a maximum of 62.5 points. Signals span January 5, 2023 to December 25, 2024.

3.3 Sector Mapping (Study 1)

Each signal's market question was classified into one of several event categories using a deterministic keyword classifier — geopolitical, political, economic, cryptocurrency, health, and other — and each category mapped to a set of sector ETFs. The mapping followed a three-level causal framework: Level 1 (direct effects, e.g., conflict → defense), Level 2 (supply-chain effects, e.g., conflict → oil disruption → airlines), and Level 3 (macro and sentiment effects, e.g., conflict → safe-haven flows → gold). Each ETF prediction carried a direction (up/down) and a confidence score from 1 (speculative) to 3 (high conviction). The final universe comprised 37 distinct ETFs spanning defense, energy, financials, real estate, clean energy, healthcare, consumer, emerging markets, safe havens, and cryptocurrency-adjacent equities. Pipeline validity was confirmed on three resolved test markets (a September 2024 Federal Reserve rate decision, the 2024 U.S. presidential election, and a Bitcoin price threshold), all of which produced sector maps matching ex ante expectations.

The distribution of observations across categories was: other (8,020), political (5,284), cryptocurrency (1,285), economic (1,280), geopolitical (853), and health (9). The health category is too small for inference and is disregarded throughout.

3.4 Equity Data

Daily equity data were obtained from Yahoo Finance via the `yfinance` library, also requiring no authentication. For each signal, I retrieved daily closing prices and computed daily log returns for all mapped ETFs and for the SPDR S&P 500 ETF (SPY), the market benchmark, over a window from 90 calendar days before to 45 calendar days after each signal. This produced 16,731 signal–ETF observations for Study 1.

4. Methodology

I employ the market-model event study (MacKinlay, 1997).

Market model. For each signal i and mapped ETF j, I estimate over a 60 trading-day estimation window preceding the signal:

R_jt = α_j + β_j · R_mt + ε_jt

where R_jt is the daily log return of ETF j and R_mt is the daily log return of SPY. Parameters are estimated by ordinary least squares.

Abnormal returns. Abnormal returns in the event window are:

AR_jt = R_jt − (α̂_j + β̂_j · R_mt)

Cumulative abnormal returns. CAR is the sum of abnormal returns over event windows of 1, 3, 7, and 30 trading days, where day 0 is the trading day on or immediately after the signal.

Cumulative average abnormal returns. CAAR is the mean CAR across all N events in a group:

CAAR(t₁,t₂) = (1/N) · Σ CAR_i(t₁,t₂)

Statistical tests. CAAR is tested against zero using a cross-sectional t-test. Directional accuracy — the share of predicted ETF movements whose sign matches the realized CAR — is tested against a 50% null using a one-sided binomial test. Study 1 and Study 2 accuracy are compared using a two-proportion z-test. All analyses are repeated by event category, confidence level, and signal magnitude.

5. Study 1 Results: Category-Level Signal Detection

5.1 The Broad Hypothesis Is Rejected

Pooling all events, up-predicted ETFs earn statistically significant negative CAAR at every horizon: −0.16% at one day, −0.28% at three days, −0.42% at seven days, and −1.06% at thirty days (all p < 0.0001; n = 12,841). Down-predicted ETFs are also negative at short horizons (−0.09%, −0.33%, and −0.52% at one, three, and seven days respectively, all significant), with the thirty-day window not significant. In other words, after a whale signal, the sectors predicted to rise fall, and the sectors predicted to fall also fall, at least initially.

This pattern is the opposite of the hypothesized sector rotation. The most plausible interpretation is that large prediction market participants are not naive directional speculators but sophisticated actors managing equity portfolios alongside their prediction-market positions. A large position may accompany hedging or liquidation of equity exposure, so the signal reflects — rather than precedes — broad equity repositioning, accompanied by mild risk-off pressure across sectors.

5.2 Finding 1: Geopolitical Signals Are Directionally Predictive

Geopolitical events behave categorically differently from the rest of the sample. Up-predicted geopolitical sectors earn a thirty-day CAAR of +1.43% (t = 2.91, p = 0.0037, n = 543) — the only positive sector CAAR anywhere in the dataset. Directional accuracy at seven days reaches 55.5% (p = 0.0008, n = 853), significantly above chance, whereas at one day it is an indistinguishable 49.2% (p = 0.68).

The combination of these results is informative: the geopolitical signal is real but operates on a lag of roughly one week, not instantaneously. This is consistent with the time required for institutional investors to reposition around geopolitical developments — fast enough that the effect exists, slow enough that it is not arbitraged away within a day. Defense (ITA, XAR), energy (XOP, XOM), and safe-haven assets (GLD, TLT) drive the effect, all through well-established economic channels.

5.3 Finding 2: Cryptocurrency Signals Are Predictive in Reverse

The cryptocurrency category produces the largest effect in the dataset, and it runs opposite to intuition. When Bitcoin or cryptocurrency probability surges, the equities mapped as beneficiaries (MSTR, COIN, MARA, RIOT) earn strongly negative CAAR: −1.58% at one day, −3.32% at three days, −6.04% at seven days, and −16.46% at thirty days (all p < 0.0001; n = 1,028). Directional accuracy as stated is 37.2% at seven days — which means accuracy of 62.8% if the prediction is reversed. The 95% confidence interval for the reverse-trade accuracy is [60.2%, 65.4%], and for the thirty-day CAAR is [−18.65%, −14.27%]; both intervals exclude the null, confirming the effect is unlikely to be chance.

This reversal is not consistent with uninformed speculation. The interpretation I advance, drawing on Kyle (1985) and Easley and O'Hara (1987), is that sophisticated participants holding long positions in cryptocurrency-adjacent equities place large bullish prediction-market positions as a partial hedge while simultaneously reducing equity exposure. The prediction-market surge thus signals informed selling of crypto equities rather than informed buying. Retail observers, seeing the surge, would plausibly interpret it as bullish — exactly the misreading that allows the reversal to persist.

5.4 Finding 3: Medium Confidence Outperforms High Confidence

A counterintuitive but robust pattern appears across confidence levels. High-confidence (level 3) predictions — assigned to the most obvious first-order effects — show no significant directional accuracy (46.9% at seven days, p = 1.00). Medium-confidence (level 2) predictions are significantly accurate: 52.8% at one day (p = 0.0002) and 52.3% at seven days (p = 0.0022). Low-confidence (level 1) predictions fall below chance.

The explanation consistent with semi-strong market efficiency is that the most obvious causal predictions are already reflected in equity prices by the time a prediction-market signal fires, leaving no exploitable lag, whereas the subtler second- and third-order effects are priced more slowly. This finding directly motivates the level-based framework: the exploitable information lives in the non-obvious links, not the headline ones.

5.5 Signal Magnitude Does Not Predict Return Magnitude

The correlation between signal size (the probability shift) and the subsequent seven-day CAR is negligible (r = −0.02, p = 0.007). Although statistically significant given the large sample, the correlation explains essentially none of the variance in returns. A signal that moves a market by 40 points conveys no more about the size of the subsequent equity move than one that moves it by 10. The signal is effectively binary: it either triggers a sector effect or it does not, and magnitude adds no predictive content. Consistent with this, large signals (Δ > 10%) and small signals (7–10%) show similar — and both below-chance — directional accuracy when pooled across all categories.

5.6 Down Predictions Are More Reliable Than Up Predictions

Across the full sample, down-predicted sectors achieve higher directional accuracy than up-predicted sectors (53.3% versus 44.5% at one day). This asymmetry is consistent with the well-documented tendency for negative information to be incorporated into prices more rapidly and forcefully than positive information.

6. Study 2 Results: Entity-Level Reasoning

6.1 Motivation and Design

Study 1's sector mapping operated at the category level: any "political" event mapped to the same ETF basket regardless of which politician, which policy, or which region was involved. Study 2 tests whether finer, entity-specific reasoning improves predictions. Two structured databases were built. A policy database encoded the known policy positions of major political figures (for example, mapping a given candidate's stated energy, defense, financial, healthcare, and trade positions to specific ETFs with confidence weights). A geographic database encoded country-specific supply-chain effects (for example, Iran → Strait of Hormuz oil-transit exposure → energy up, airlines and shipping down). An entity-extraction step identified the people, countries, and organizations named in each market question and routed to the appropriate database, falling back to the Study 1 category map when no entity matched.

This produced 13,549 signal–ETF observations, computed through the identical event study pipeline so that the only difference between studies is the sector mapping.

6.2 Entity-Level Reasoning Degrades Accuracy

The central result of Study 2 is that the added specificity made predictions worse, not better, and significantly so. Pooled seven-day directional accuracy fell from 46.4% in Study 1 to 43.6% in Study 2 (two-proportion z = −4.83, p < 0.0001). At one day the decline was from 46.5% to 44.3% (z = −3.93, p = 0.0001). The degradation was concentrated in exactly the categories where Study 1 had shown the most promise: geopolitical seven-day accuracy collapsed from 55.5% to 42.7%, and political accuracy also declined.

6.3 The Exception: High-Conviction Bearish Predictions

One subgroup improved. High-confidence (level 3) bearish predictions in Study 2 achieved 62.2% directional accuracy at seven days (p = 0.0008), an improvement over the corresponding Study 1 figure. This is the single case in which entity-specific reasoning added value: when the model is highly confident that a specific entity's circumstances will harm a specific sector, the added specificity helps. Everywhere else, it hurts.

6.4 Interpretation: Categorical Versus Entity-Specific Information

The degradation is itself the finding. The most coherent interpretation is that prediction markets transmit broad categorical uncertainty to equity markets — "something geopolitical is escalating" — rather than entity-specific detail. The entity-specific implications ("this particular conflict threatens this particular shipping lane") are precisely the kind of information that analysts, institutions, and existing markets already track and price. By reasoning at the entity level, Study 2 effectively bet on information that was already incorporated into prices, introducing noise rather than signal. The lone exception — high-conviction bearish entity predictions — is consistent with this view: only when the entity-specific implication is both severe and underappreciated does the added detail pay. In effect, the contrast between the two studies localizes the exploitable information to the broad, categorical layer of the signal.

7. Discussion

7.1 A Unified Picture

Taken together, the two studies suggest that prediction market whale signals are not uniformly informative for equities, but contain specific, locatable pockets of information. The broad signal is mildly risk-off, consistent with hedging. Geopolitical signals carry genuine directional content on a one-week lag. Cryptocurrency signals are informative in reverse, consistent with informed equity selling disguised as bullish prediction-market activity. And the information that is exploitable is categorical and second-order, not entity-specific and obvious — which is why Study 2's added specificity backfired.

7.2 Trading Implications

Two strategies follow from the findings. The first is a geopolitical strategy: on detection of a geopolitical whale signal, take long positions in defense and safe-haven ETFs (ITA, GLD) held for seven days. With 55.5% directional accuracy this is a thin but positive edge; because the abnormal return (+1.43% over thirty days) is small relative to transaction costs, this strategy is best expressed through low-cost ETFs rather than options. The second is a cryptocurrency strategy: on detection of a large bullish Bitcoin signal, take bearish positions in crypto-adjacent equities (for example, put options on MSTR) held for thirty days. With reverse accuracy of 62.8% and a large effect size, this strategy can plausibly survive option premiums.

A simple power calculation illustrates the practical requirement. The individual-trade standard deviation of crypto CARs is approximately 36%, so single trades are highly volatile and prove nothing in isolation. However, the standard error of the mean shrinks as the square root of the number of trades. Requiring the upper bound of the 95% confidence interval on the thirty-day return to fall below an assumed 1.5% option cost implies a break-even sample of roughly 23 independent trades. The critical caveat is independence: clustered signals during a single market regime are correlated and yield less statistical power, so trades must be spaced and position-limited. Overall portfolio variance is best reduced not by accumulating correlated crypto trades but by combining the crypto and geopolitical strategies, which are largely uncorrelated — the standard logic of the Sharpe ratio.

7.3 Market Efficiency Implications

The results are consistent with semi-strong-form efficiency operating selectively. Obvious, widely-tracked causal chains (the high-confidence, entity-specific predictions) are priced quickly; subtle, second-order effects (the medium-confidence, categorical predictions) are priced more slowly. The seven-day geopolitical lag and the medium-confidence outperformance both point to the same conclusion: the market is efficient with respect to information that is easy to reason about, and less efficient with respect to information that requires effort to connect. The cryptocurrency reversal additionally illustrates how the interpretation of a public signal, not merely its existence, governs whether it is exploitable.

7.4 Study 3: Live Forward Validation (Ongoing)

Because Study 1 and Study 2 are retrospective, a live out-of-sample test is underway. A monitoring application polls Polymarket every five minutes, detects whale signals in real time, records the relevant equity prices at detection, and re-checks those prices at seven and thirty days to compute realized directional accuracy against the 62.8% backtested cryptocurrency baseline. This forward test, beginning June 2026, will determine whether the documented edge persists as prediction markets grow and as awareness of such signals spreads. Code and the live tracker are publicly available.

8. Limitations

Several limitations constrain interpretation. The analysis is retrospective until Study 3 concludes; past performance need not persist, particularly as prediction-market volume grows and any edge is competed away. The health category (n = 9) is too small for inference and is excluded. The keyword classifier infers categories from question text rather than from verified metadata, and markets matching no keyword received a generic basket, which may dilute signal quality for novel event types. Transaction costs are estimated rather than measured, and the trading implications in Section 7.2 should be read as upper bounds on achievable returns. The independence assumption underlying the break-even calculation may be violated when signals cluster within a market regime. Finally, the cryptocurrency reversal interpretation, while consistent with informed-trading theory, is inferred rather than directly observed; confirming it would require participant-level position data that Polymarket does not publicly disclose.

9. Conclusion

This paper provides the first large-sample event study of prediction market whale signals on equity sector returns, using 2,623 signal events across 604 resolved Polymarket markets from 2023 to 2024. The broad hypothesis — that whale signals predict positive sector rotation — is rejected; signals instead precede mild, broadly negative abnormal returns, consistent with hedging behavior by sophisticated participants. Three specific findings survive: geopolitical signals are directionally predictive at a seven-day horizon (55.5%, p = 0.0008); cryptocurrency signals are predictive in reverse, with predicted-beneficiary equities falling 16.46% over thirty days (p < 0.0001); and medium-confidence second-order predictions outperform high-confidence first-order predictions. A second study found that entity-level reasoning degraded accuracy across the board, implying that prediction markets transmit broad categorical uncertainty to equities rather than entity-specific detail, which is already priced. These results locate specific, theoretically coherent pockets of exploitable information within an otherwise efficient market, and they suggest that the value of prediction-market data for equity trading lies not in granular reasoning but in the broad, fast detection of categorical signals — most notably the counterintuitive cryptocurrency reversal. A live forward-validation study is ongoing. All data and code are public and the full pipeline is reproducible from freely available sources.

References

1. Ball, R., & Brown, P. (1968). An empirical evaluation of accounting income numbers. Journal of Accounting Research, 6(2), 159–178.
2. Berg, J., Nelson, F., & Rietz, T. (2008). Prediction market accuracy in the long run. International Journal of Forecasting, 24(2), 285–300.
3. Chen, H., De, P., Hu, Y., & Hwang, B. H. (2019). Wisdom of crowds: The value of stock opinions transmitted through social media. Review of Financial Studies, 27(5), 1367–1403.
4. Da, Z., Engelberg, J., & Gao, P. (2011). In search of attention. Journal of Finance, 66(5), 1461–1499.
5. Easley, D., & O'Hara, M. (1987). Price, trade size, and information in securities markets. Journal of Financial Economics, 19(1), 69–90.
6. Fama, E. F. (1970). Efficient capital markets: A review of theory and empirical work. Journal of Finance, 25(2), 383–417.
7. Fama, E. F., Fisher, L., Jensen, M. C., & Roll, R. (1969). The adjustment of stock prices to new information. International Economic Review, 10(1), 1–21.
8. Gürkaynak, R. S., Sack, B., & Swanson, E. (2005). The sensitivity of long-term interest rates to economic news. American Economic Review, 95(1), 425–436.
9. Hanson, R. (2003). Combinatorial information market design. Information Systems Frontiers, 5(1), 107–119.
10. Hayek, F. A. (1945). The use of knowledge in society. American Economic Review, 35(4), 519–530.
11. Jensen, M. C., & Ruback, R. S. (1983). The market for corporate control: The scientific evidence. Journal of Financial Economics, 11(1–4), 5–50.
12. Katona, Z., Painter, M., Patatoukas, P. N., & Zeng, J. (2018). On the capital market consequences of alternative data: Evidence from outer space. Working paper, University of California, Berkeley.
13. Kyle, A. S. (1985). Continuous auctions and insider trading. Econometrica, 53(6), 1315–1335.
14. MacKinlay, A. C. (1997). Event studies in economics and finance. Journal of Economic Literature, 35(1), 13–39.
15. Preis, T., Moat, H. S., & Stanley, H. E. (2013). Quantifying trading behavior in financial markets using Google Trends. Scientific Reports, 3, 1684.
16. Wolfers, J., & Zitzewitz, E. (2004). Prediction markets. Journal of Economic Perspectives, 18(2), 107–126.

Appendix A: Summary of Key Statistics

Appendix B: Data and Code Availability

All data were sourced from public APIs requiring no authentication (Polymarket Gamma and CLOB APIs; Yahoo Finance via `yfinance`). The complete analysis pipeline — data collection, signal detection, sector mapping, event study computation, statistical testing, and visualization — is available at: `github.com/naterjlevy-a11y/polymarket-whale-signals`. The pipeline is fully reproducible: install dependencies from `requirements.txt` and run the modules in sequence.