Research Report: KXBTC15M Late-Conviction Momentum Screening

Short Disclaimer

This is a historical simulation study only. Nothing here is investment advice or a claim about future performance.

Intro / Thesis

This report covers a parameter sweep on a custom Kalshi strategy trading KXBTC15M contracts. The core idea: when both 5-minute and 15-minute Coinbase BTC spot momentum agree with an already-favored 80–95 cent contract in the final 1–3 minutes before settlement, holding that contract to settlement may produce positive net P&L after fees.

The single-screen result of +$51.84 net P&L (42 trades, 95.2% win, $35.48 max drawdown, Sharpe 0.25 over Aug 6–Sep 5) was the starting point. That result was selected from a broader search, so it is treated as a screening outcome, not proof of edge. The sweep across 100 completed variants tells a more complete and less flattering story.

Variant and Strategy Explanation

The base strategy is a custom Kalshi DSL that runs on a 10-second loop. It selects the most liquid KXBTC15M market tied to BTC, pulls Coinbase BTC-USD 5-minute and 15-minute change fields every 10 seconds, and applies two entry rules in the final 60–180 seconds before expiry:

• Buy YES when 5m change ≥ +0.0008, 15m change ≥ +0.0015, spread ≤ 0.02, and yes_best_ask is between 0.80 and 0.95.
• Buy NO when the same thresholds are flipped negative, with no_best_ask between 0.80 and 0.95.

Position size is 40 per entry, with a max position of 200 and one entry per market. A price floor of 0.05 and ceiling of 0.95 are applied as risk guardrails, and a standing cancel_all rule clears resting orders.

The parameter sweep varied only risk.price_floor and risk.price_ceiling. Floor values ranged from 0.05 to 0.45 in roughly 0.04–0.05 increments. Ceiling values ranged from 0.55 to 0.95. That's a 10×10 grid, 100 cells total, all of which completed successfully. Each successful variant is saved as a runnable Turbine strategy.

The key thing to understand: most of the sweep was flat or near-flat. Only variants with a ceiling at or above roughly 0.82 generated meaningful trade counts and P&L. Lower ceilings either produced zero trades or a small losing sample.

Top Results

The top eight variants are all functionally identical in outcome:

All eight share a price ceiling of 0.95. The floor varies from 0.05 up to 0.36 but doesn't change the result. That's a signal in itself: the floor parameter is doing almost no work in this region. The ceiling is what matters, and it matters because it determines whether the strategy is allowed to trade at all.

The naive view of these results is that they look good: ~19% ROI, 94% win rate, mild drawdown. But that interpretation ignores how they were found and how fragile the edge is under scrutiny.

The robustness statistics tell the real story here. The deflated Sharpe for the winner is 0.23, while the expected maximum Sharpe across the sweep is 0.36. That's worse than the expected best of a skill-less sweep. In plain English: the best result from this parameter search is indistinguishable from what you'd get by picking the luckiest cell from random noise.

The permutation test reinforces this. It shuffled the edge-feed values against trade prices — specifically the Coinbase momentum feed timing — and reran the sweep 100 times against time-scrambled versions of the same feed. The real best net P&L was $38.13. The permutation test produced an upper p-value of 0.248.

That means the observed edge-feed timing beat 75.2% of the time-scrambled versions. It did not clear the 95% bar that would suggest the feed timing is doing meaningful work beyond chance. The strategy's price-based conditions were not permuted, so this test only challenges the edge-feed component. Still, the headline is plain: p = 0.248.

Additional warning flag: 47% of fills in the winner cell were at prices below 0.10 or above 0.90. Those are the price ranges where fill assumptions in simulation are least trustworthy. Real order books at those levels may not fill the same way a backfill assumes they do.

None of the top variants should be described as strong, validated, or promising. They are the best of a 100-cell sweep, and the statistical evidence says that's consistent with selection noise.

Bottom Results

The bottom of the sweep is dominated by low price ceilings that prevent the strategy from trading at all. Several cells show zero trades, zero P&L, and zero drawdown — not because the strategy lost, but because it never entered a position.

The only actual losing variant in the bottom grouping:

Two trades, both losers. The sample is tiny, so this isn't evidence of a real short-side edge either. It just shows what happens when a low ceiling restricts the strategy to a handful of marginal setups.

The rest of the bottom cells (ceilings 0.59, 0.64, 0.68, 0.73, 0.77, and 0.82 with various floors) all show zero trades. The marginal ceiling analysis confirms this pattern: mean P&L for a 0.55 ceiling is -$1.55, then flat at zero through 0.77, before jumping at 0.82 and above. The edge, such as it is, lives entirely in the high-ceiling region — which is also where fill assumptions are weakest.

Conclusion

The raw result of +$51.84 net P&L on the single-screen variant initially looked interesting. The full sweep does not support it.

Here's the honest summary:

• The strategy only trades meaningfully when the price ceiling is ≥ 0.82, and the 0.95 ceiling produces the best raw P&L.
• The deflated Sharpe (0.23) is below the expected maximum Sharpe for a skill-less sweep (0.36).
• The permutation test returns p = 0.248, meaning the edge-feed timing is not distinguishable from chance.
• 47% of simulated fills occur at extreme prices below 0.10 or above 0.90, where fill assumptions are least reliable.
• The parameter sweep mostly produces flat or no-trade results. The positive P&L is concentrated in a narrow corner of the grid, which is exactly what overfitting looks like.

This is not a strategy I would put capital behind without independent out-of-sample testing and live-market fill verification. The top results are consistent with selection noise, not a discovered edge.

Long Disclaimer

This document is a historical simulation research report prepared for internal analytical purposes. It describes the results of a parameter sweep over a custom automated trading strategy expressed in a domain-specific language and executed against historical market data from Kalshi and Coinbase.

Historical simulation results, including net P&L, ROI, Sharpe ratio, maximum drawdown, win rate, and trade counts, are not indicative of future performance. They are based on assumptions about order fills, spreads, fees, market liquidity, and data alignment that may not hold in live trading. In particular, simulated fills at prices below $0.10 or above $0.90 are flagged in this report as less trustworthy due to known execution risks in thin order books.

The statistical robustness measures included here — deflated Sharpe ratio, marginal parameter analysis, neighborhood degradation, and permutation testing — are tools for assessing whether a result may be attributable to overfitting or selection bias. They are not guarantees of live performance. A deflated Sharpe below the expected maximum for a null sweep, or a permutation p-value above 0.05, indicates that the result is consistent with chance and should not be interpreted as a validated edge.

This report does not constitute investment advice, a recommendation to buy or sell any security, derivative, prediction market contract, or cryptocurrency. It does not imply any future profit or predict any market movement. Trading prediction market contracts involves substantial risk, including the risk of total loss of the amount traded.

Each successful variant referenced in the sweep is saved as a runnable Turbine strategy for further analysis and reproducibility. Saving a strategy does not constitute an endorsement of that strategy for live trading.

Anyone reading this report should perform their own due diligence, understand the relevant market rules and fee schedules, and consult a qualified financial professional before making any trading decisions.