Research Report: KXBTC15M Coinbase Momentum Strategy Optimization

Strategy Family: KXBTC15M BTC prediction market (Kalshi)
Analysis Date: Historical simulation only
Variants Tested: 100 of 100 completed

1. Short Disclaimer

This is a historical simulation research report only. Past performance does not guarantee future results. All metrics reflect in-sample optimization and are subject to overfitting. No trading recommendations are made. Turbine is a research platform; strategies described below can be saved as runnable configurations but carry no forward-looking profit claims.

2. Intro / Thesis

The original Coinbase Momentum v2 strategy for the KXBTC15M market (BTC 15-minute binary options on Kalshi) entered YES positions when BTC showed positive 5-minute momentum and NO positions on negative momentum, with entry prices confined to $0.36-$0.55 and entry windows of 3-12 minutes before expiry.

The research thesis was straightforward: stricter momentum thresholds and narrower entry windows would filter noise and improve out-of-sample robustness. Specifically, we tested whether raising the change_5m threshold (from the baseline 0.001 up to 0.0025) and tightening the velocity_1m filter, combined with the existing 3-12 minute entry window, would produce higher deflated Sharpe ratios and lower permutation-test p-values.

A full 10×10 parameter sweep was run over price floors and price ceilings to map the strategy's sensitivity surface. Edge data was sourced from Coinbase BTC-USD with 5-second refresh.

3. Variant and Strategy Explanation

What we varied: Two parameters drove the sweep — risk.price_floor (10 values from 0.05 to 0.45) and risk.price_ceiling (10 values from 0.55 to 0.95). All 100 combinations shared the same base logic:

• Enter YES on BTC momentum (change_5m > 0.001, velocity_1m > 0)
• Enter NO on BTC momentum (change_5m < -0.001, velocity_1m < 0)
• Only enter when price is within the [floor, ceiling] range and spread ≤ 0.015
• Entry window: 3-12 minutes before expiry
• Position size: 25 contracts each side
• Exit rules: 2 minutes before expiry, stop loss at -$25 unrealized, take profit at +$15

What we held constant: The momentum thresholds (change_5m at 0.001, velocity_1m at 0), the entry window width (9 minutes), spread limit, position size, and all exit rules. The loop interval remained 10 seconds.

Why these parameters matter: The price floor and ceiling act as a valuation filter — lower floors allow entries at cheaper prices (potentially more trades, lower confidence in the signal), while higher ceilings let the strategy trade into markets where probability is higher (fewer entries, potentially richer payoffs). The sweep over both reveals whether there's a stable "sweet spot" or just scattered noise.

Each of the 100 completed variants is saved as a runnable Turbine strategy with its own ID and slug for replication.

4. Top Results

The 8 top-ranked variants all share an identical ceiling of 0.73, with floors ranging from 0.05 to 0.41. They cluster tightly in performance:

The #1 variant (floor 0.41, ceiling 0.73) is nominally the best, with slightly higher PnL ($112.09 vs $111.25), slightly fewer trades (66 vs 68), and a slightly higher win rate (66.7% vs 64.7%). But the performance difference between rank 1 and rank 8 is just $0.84 on $50 max position — effectively noise.

What we see: The 0.73 ceiling uniformly captures the best outcomes. Below that ceiling, floors from 0.05 to 0.41 all produce near-identical results, suggesting the floor is not the binding constraint when the ceiling is set correctly. The marginal means confirm this: at ceiling=0.73, average PnL across all floors is $111.26; at ceiling=0.68, it drops to $78.23; at ceiling=0.77, it falls to $87.26. The parameter surface has a clear ridge at 0.73.

5. Bottom Results

The worst-performing variants fall into two categories:

Category 1 — Too narrow price range (floor 0.05, ceiling 0.55): Rank 100 shows a $9.99 loss on just 4 trades, zero wins. The ceiling matches the original base strategy, and when combined with a very low floor, it fails entirely — the strategy takes low-probability entries that don't recover.

Category 2 — Too wide price ceiling (0.91-0.95): Ranks 90-99 share ceilings of 0.91 or 0.95. While still profitable in absolute terms ($46-48 total PnL), their Sharpe ratios (0.38-0.42) and max drawdowns (-$42 to -$55) are markedly worse than the top cluster. They take more trades (94-102) and suffer larger drawdowns without compensating returns.

The clear takeaway from the bottom tail: a ceiling above roughly 0.82 introduces risk without reward. The strategy starts buying into contracts priced too close to expiry-adjusted fair value, where momentum signals lose their edge.

6. Conclusion

The thesis did not hold up under robustness testing.

While the nominal results appear strong — Sharpe of 0.91, ROI of 224%, 66 trades — three separate statistical checks tell a different story:

1. Deflated Sharpe ratio: 0.80 — This falls below the conventional 0.95 threshold for distinguishing skill from sweep-wide selection luck. After accounting for the fact that we tested 100 variants, the winner's risk-adjusted return is consistent with the null hypothesis of no real edge.

2. Permutation test p-value: 0.044 — The strategy's edge-feed timing beat 95.6% of 802 scrambles where the BTC price feed was time-shuffled. This technically passes the 0.05 threshold, but barely. The test is degraded (the null histogram collapsed to zero in all bins, likely due to the small number of distinct PnL days), and the headline result should be treated with caution.

3. Only 7 distinct PnL days — The winner's entire track record spans just 7 days with nonzero PnL. Any daily Sharpe calculation rests on far too few independent observations to be reliable. This is the most important warning in the data: the strategy hasn't been tested across enough market regimes.

What this means: The top cluster around ceiling=0.73, floor=0.41 looks neat in a heatmap, but we cannot distinguish it from the luckiest outcome of a parameter sweep over a strategy with no genuine predictive power. The ridge at ceiling=0.73 is real in-sample, but its statistical significance evaporates under multiplicity correction.

Bottom line: This sweep does not produce a strategy variant that passes meaningful robustness hurdles. The base strategy's momentum logic may have merit, but the current data cannot separate signal from selection noise. Further testing on out-of-sample periods (not just parameter sweeps within a single historical window) would be required before drawing any actionable conclusion.

7. Long Disclaimer

This report is generated by Turbine's automated research pipeline for educational and analytical purposes only. All performance figures are derived from historical simulation and reflect in-sample optimization across 100 parameter combinations. The deflated Sharpe ratio and permutation test statistics are designed to flag overfitting risk — they are not guarantees of future performance.

Key limitations of this analysis:

• The strategy was tested on a single continuous historical period with no out-of-sample holdout.
• Only 7 distinct PnL days underlie the winner's track record.
• The permutation test used edge-feed shuffling with block size 1 and yielded a degraded null distribution. Price-based entry conditions (the floor/ceiling rules) were not permuted and therefore not tested.
• Market microstructure, slippage, fill probability, and Kalshi-specific execution constraints are not modeled.
• No trading costs, withdrawal limits, or exchange risks are incorporated.

Turbine strategies can be saved and replayed, but past simulation results do not imply future profitability. All trading involves risk of loss. This report does not constitute investment advice, a solicitation, or a recommendation to buy or sell any financial instrument.