F2 Anticipator V2 — Can a late-book signal anticipate 5-minute BTC Up/Down markets?

Strategy summary

F2 Anticipator V2 trades Polymarket's BTC Up/Down 5-minute markets. At T+150 seconds, it estimates the probability that Up will settle.

That estimate is built exclusively from causal, book-only observations captured at T+90, T+120, T+135 and T+150: price, spread, displayed depth and order-book imbalance for both sides.

It buys the side with the largest estimated value, and only when the expected edge is at least 2 cents per share after taker fees.

The model, its features and the 2-cent threshold were frozen before the strict post-freeze evaluation window began.

This is a research replay of historical PMData order-book snapshots, not a live trading bot: it does not place orders.

Economic hypothesis

The frozen hypothesis: between T+90 and T+150, changes in the BTC binary order book may carry enough information to estimate the settlement probability before the market's conventional T+180 entry point. A deliberately low-capacity logistic model should be able to exploit temporary mispricing — but only when its estimated value exceeds the visible ask plus the dynamic taker fee.

Before any OOS evaluation, six survival criteria were frozen and all had to hold: at least 300 signals across at least 15 qualified days; positive average EV; a day-level bootstrap 95% lower bound above zero; positive EV in both temporal halves of the sample; positive EV under a 1-cent execution stress; and a model Brier score below the raw order book’s own Brier score.

Retuning the model, features or threshold on the same OOS sample after seeing these results would create a new, contaminated hypothesis — the frozen V2 hypothesis is treated as closed on this sample regardless of the outcome.

Signal and timing convention

The causal timing convention is explicit and checked by the model-integrity self-test in the repository:

• Bilateral book snapshots are captured at T+90, T+120, T+135 and T+150 — strictly before the trade decision, never after.
• The frozen logistic artifact estimates P(Up wins) directly from those causal snapshots.
• Net edge is computed for both Up and Down using a dynamic taker-fee model.
• The strategy trades the better side only when edge is at least 2 cents, the ask sum is between 0.98 and 1.04, entry price is between 0.05 and 0.90, and displayed ask depth is at least five shares.

p_up = model(book_90, book_120, book_135, book_150)
fee(p) = 0.07 * p * (1 - p)
edge_up   = p_up       - ask_up   - fee(p_up)
edge_down = (1 - p_up) - ask_down - fee(1 - p_up)
trade side = argmax(edge_up, edge_down)
enter only if max(edge_up, edge_down) >= 0.02

Temporal OOS protocol

• Evaluation period: 7 June 2026 to 10 July 2026 — 34 calendar days.
• 29 complete, qualified trading days out of those 34.
• 8,352 markets in the evaluation window; 8,177 valid feature rows survived the causal book-freshness checks.
• 2,007 quote-level signals met the edge, price-band and depth conditions.
• Confidence intervals use a day-level bootstrap with 10,000 resamples.
• No book snapshot older than 2 seconds is accepted at any decision timestamp.
• The taker fee model is 0.07 × p × (1 − p), applied dynamically to the model’s own probability estimate.
• A quote-persistence test requires the same or a better executable quote to remain available one second later.
• Raw licensed PMData is excluded from this repository and this page; only aggregate, non-proprietary outputs are published.

Results

• Average predicted edge was +3.31c, but realized quote-level EV was only +0.04c.
• The model's Brier score was 0.165898, against 0.165569 for the raw order book — the model did not improve probability calibration relative to the market.
• Higher predicted-edge quartiles did not produce higher realized EV.
• Quote persistence deteriorated among the supposedly strongest signals.
• The later half of the monthly sample and the strict post-freeze sample were both negative.

Charts

Why the strategy failed

A NO-GO here is a successful validation outcome, not an embarrassment: the protocol did exactly what it was built to do — reject a plausible-looking edge that does not survive realistic scrutiny.

• The apparent model edge was not calibrated against realized outcomes.
• The result was highly sensitive to quote persistence and execution stress.
• Performance deteriorated after the model freeze.
• The strongest predicted signals did not produce the strongest realized returns.
• A marginally positive headline estimate with a confidence interval crossing zero was insufficient.
• The persisted and strict post-freeze results were negative.

What this case demonstrates

• Predicted edge must be checked against a realized-outcome benchmark, such as book-relative Brier score, before it is treated as informative.
• A quote-persistence and execution-stress test can invalidate an average-EV result that looks acceptable on paper.
• A strict post-freeze evaluation window is a meaningful additional control beyond a monthly temporal OOS.
• Segmenting by predicted-edge quartile can reveal that a model’s most confident signals are its weakest, not its strongest.
• Freezing the model, features and threshold in advance is what makes a NO-GO a real, uncontaminated result rather than an invitation to keep retuning.

Reproducibility

• Frozen logistic-model artifact with a SHA-256 manifest binding the model to its spec.
• The exact feature list, fee model and execution contract are committed in the repository.
• A causal, resumable PMData replay script verifies the frozen model hash before scoring.
• Aggregate, machine-readable JSON and CSV results are committed alongside the code.
• A reproducible chart-generation script regenerates both committed charts from the aggregate results.
• Unit tests and a model-integrity self-test verify the frozen artifact before every replay.
• Raw PMData files and market-level derived rows are intentionally excluded: the source dataset is licensed and cannot be redistributed.

View source on GitHub (opens in a new tab)

Limitations

• Only 29 complete, qualified days were available in the primary OOS window.
• The 34-day block was temporal OOS relative to training, but four June days had already been viewed during V2 development.
• The cleaner strict post-freeze slice contained only 12 complete days.
• Confidence intervals remain wide relative to the estimated effect size.
• Quote persistence is only a proxy for fillability, not a complete queue-position or market-impact model.
• Results apply to the tested BTC 5-minute Polymarket sample, not necessarily to other markets or timeframes.
• PMData's licensed raw files cannot be independently redistributed, which limits third-party re-verification to the aggregate outputs published here.
• This is historical research, not investment advice or a forecast.

Formal verdict

Formal validation summary

Deployment: NOT TESTED

A PASS on causal timing or reproducibility only means the experiment was built correctly — it is not evidence that the strategy is profitable. Correct implementation cannot rescue a claim that fails on calibration and economic viability.

Formal verdict: NO-GO

NO-GO due to economic and temporal-stability failure. The signal did not demonstrate incremental, monetizable edge under quote persistence, execution stress and a strictly post-freeze evaluation. Progression to shadowing is not recommended without a new hypothesis and a fully independent validation.