This repository publishes reproducible research results—including negative results—on BTC 5-minute Up/Down markets on Polymarket. It is not a trading bot, an investment signal, or a recommendation to trade.
Across the hypotheses evaluated so far, we have not found a durable, economically executable directional advantage over the Polymarket order book in BTC 5m markets. Some signals predict the underlying asset, but the recurring result is that the market price already incorporates that information more effectively, or that fees, the ask price, non-fills, and adverse selection consume the gross edge.
This does not prove that Polymarket is efficient for every asset, horizon, or regime. The conclusion is limited to the data, windows, entry rules, and costs documented here. Confidence intervals crossing zero are classified as inconclusive, not as proof of no effect.
| Research line | OOS sample | Main result | Verdict |
|---|---|---|---|
| Theoretical fair-value repricer | 7,506 markets, 27 days | AUC 0.8355 versus 0.8600 for the book; net EV -1.83 c/share, 95% CI [-2.95, -0.71] | Rejected |
| IBS, Binance flow, and volatility regime on top of the book | 4,649 markets, 17 days | Brier worsens in all five chronological blocks; the primary threshold loses -3.552 c/share | Rejected |
| IBS + RSI + Fibonacci + HAR-GASVR | 16,310 markets | Net EV -2.139 c/share; 95% CI [-2.927, -1.267]; 0/5 positive folds | Rejected |
| Digital residual versus the book | 7,506 markets, 27 days | The model has directional information, but the book has better AUC and Brier | Not executable |
| F2 digital-value | 1,391 signals, 27 days | EV -0.44 c/share; 95% CI [-2.82, +1.84] | Inconclusive; not promotable |
| F2 execution audit | 119 intended orders | Observed price improvement is offset by non-fills and adverse selection | Closed under the frozen rule |
| WebSocket lead-lag | 1,582 signals, 18 days | EV per signal -2.646 c/share; 95% CI [-4.558, -1.048] | Rejected |
| Volatility HMM | 19,300 opportunities | No variant with a positive 95% net-EV confidence interval | Not confirmed |
| Up/Down parity | 5,176,651 snapshots, 70 days | No sub-unit ask and no gross edge before fees | Not executable |
| F2 late favorite, L2 replay | 440 taker signals; 380 maker signals | EV -4.0 c/contract taker and -4.3 c/contract maker; adverse selection measured directly | Closed |
| Book-and-tape microstructure V3 | 4,050 observations, 11 folds | Market wins Brier in 10/11 folds; one extra tick makes all simulated EV negative | Rejected |
| Causal noise / first-passage | 2,733 trades per rule | Net EV -4.76 c (MOM) / -4.78 c (REV), with 5/5 negative folds | Rejected |
| Extreme-markout ranker | 172 OOS selections | Real gross lift, but net EV -0.37 c/share and 95% CI crosses zero | Inconclusive |
The complete reports are in research/. Inconclusive and negative results are included to avoid retrospectively selecting only favourable outcomes.
The primary published matrix is data/sii_exec_matrix_btc5m.parquet: 23,338 BTC 5m markets from 2026-02-12 to 2026-05-04 UTC. It contains binary label_up labels and point-in-time observed prices at 30, 60, 120, 180, and 240 seconds after each market started. Its SHA-256 is:
FAD173C57D0A8A57050F84E9A826820F2667F908949019B06C97E27773EC9673
The matrix was derived from trades.parquet and markets.parquet in the SII-WANGZJ/Polymarket_data dataset, which documents on-chain OrderFilled events and is MIT licensed. Full attribution is in THIRD_PARTY_NOTICES.md; reconstruction is described in docs/DATA.md.
Other sources used by individual research lines, but not redistributed here, include:
- The public Polymarket Gamma API, for metadata and resolution checks.
- One-minute BTCUSDT candles from the Binance USD-M Futures API, causally transformed to 5-minute features when the hypothesis requires underlying-market variables.
- The historical order book from kachoio/polymarket-5-minute-crypto-up-down-markets, under CC0-1.0, for execution tests requiring bid/ask and depth.
We do not include account data, private orders, keys, execution services, deployed models, or production artifacts.
- The hypothesis, primary test cell, and thresholds are frozen before testing.
- Temporal walk-forward validation; future observations are never mixed into training or calibration.
- Official or on-chain labels where available; inferred outcomes are not used to determine results.
- Executable prices and dynamic taker fees when applicable; proxy prices are labelled as optimistic.
- Day-level bootstrap, temporal-half checks, and explicit OOS sample sizes.
- A 95% confidence interval crossing zero is inconclusive. A well-identified negative result is a no-go.
The published matrix can be inspected with Python 3.11+ and pyarrow:
import pyarrow.parquet as pq
matrix = pq.read_table("data/sii_exec_matrix_btc5m.parquet")
print(matrix.schema)
print(matrix.num_rows)Repeating a study requires the external sources and exact configuration described in its report. We publish results and small derived data that support auditing, rather than promising instant reproduction of a historical pipeline that can require hundreds of GB.
data/ Derived SII matrix and column dictionary
docs/ Methodology, sources, and limitations
research/ Report for each hypothesis, including no-go and inconclusive results
Original repository material is released under the MIT License. Derived data retain the attribution obligations of their source; see THIRD_PARTY_NOTICES.md.