Turbine Research Report — Kalshi KXBTC15M

Short Disclaimer

Historical simulation only. No live trading results. No guarantee of future performance.

Intro / Thesis

This report covers the third validation of the "M5 remote-plane" idea: using multi-horizon BTC trend confirmation to predict repricing in Kalshi's KXBTC15M 15-minute BTC markets.

The hypothesis was simple — when both short-term (5-minute) and medium-term (1-hour) BTC returns are positive, buy YES. When both are negative, buy NO. Combine that with a price band of $0.25–$0.75, a spread cap of $0.03, and a minimum time-to-expiry of 2 minutes. The goal was to test whether trend agreement produces positive, repeatable historical performance across a 30-minute permutation budget of 100 completed variants.

Spoiler: it doesn't.

Variant and Strategy Explanation

The strategy is a custom Kalshi loop running every 30 seconds against the most liquid KXBTC15M contract. It pulls Coinbase BTC-USD data refreshing every 10 seconds, specifically change_5m and change_1h.

Entry rules:

• Buy YES when both BTC changes are positive and price sits between $0.35 and $0.65 with spread ≤ $0.03 and >2m to expiry.
• Buy NO when both BTC changes are negative and price sits at exactly $0.65 (a DSL quirk — both the low and high bound for the NO side are set to 0.65 in the base configuration) with spread ≤ $0.03 and >2m to expiry.

Exits are a hard stop at −$25 unrealized PnL and a flatten rule at 2 minutes before expiry. Position sizing is 1 contract per signal, max position 100, max loss $25.

I ran 100 completed variants by sweeping entry band lower bounds from $0.25 to $0.45 and upper bounds from $0.55 to $0.75 in steps of $0.05 (25 parameter cells, each with spread cap permutations).

Top Results

The "best" variants all produced identical outcomes:

The top 8 variants — essentially every combination of entry bands from $0.25–$0.55 through $0.25–$0.65 with spread caps of $0.01 to $0.04 — all landed at the exact same −$9.02 PnL. That's a red flag on its own: the entry band parameters were not actually exercising different behavior. The sweep varied entry_low and entry_high, but every cell produced identical trades. These axes were inert within the tested window.

A 51.2% win rate on 4,281 trades sounds close to break even, until you account for fees. The robustness check flagged that fees consumed 111% of the winner's gross PnL — meaning the strategy was net negative before fees were even fully applied.

The deflated Sharpe for the top variant is 0.35, well below the 0.95 threshold needed to distinguish it from the luckiest result of a skill-less sweep. The raw Sharpe of −0.07 is negative anyway.

The permutation test (edge-feed scramble, n=1,000) returned a p-value of 0.001, meaning the timing of the edge feed beat 99.9% of time-scrambled permutations. However — and this is important — the market price series was not permuted. Price-based entry conditions were not tested by this design. In context with the inert axes and negative net PnL, this p-value doesn't rescue the strategy.

Every top variant has been saved as a runnable Turbine strategy, but none of them warrant deployment based on this data.

Bottom Results

The worst variant of the 100 was the band $0.25–$0.55 with spread cap $0.01:

The spread cap of $0.01 was the only parameter that produced a meaningfully different outcome — by trimming the trade count to 3,953 and dropping the win rate to 46.6%, it made things slightly worse and doubled the drawdown severity.

The bottom of the table reinforces the top: the entire parameter surface sits in a narrow negative band of −$9.02 to −$10.14. There is no green zone hiding anywhere in this sweep.

Conclusion

This variant family fails validation.

Multi-horizon BTC trend confirmation, as designed here, does not produce positive historical performance in KXBTC15M. The best outcome across 25 parameter cells and 100 variants was a −9.02% ROI with negative Sharpe and max drawdown exceeding total PnL.

The robustness statistics are unambiguous:

1. Deflated Sharpe of 0.35 — consistent with selection noise. This is not a robust edge.
2. Fees consume 111% of gross PnL — even the raw signal, before costs, was not enough to survive.
3. Both entry parameters were inert — the sweep did not actually explore different behavior. Flatness here is not robustness; it's a sign the conditions were never triggered differentially.
4. Permutation p-value of 0.001 — superficially interesting, but the price series itself was not scrambled, and the net result is still negative.

The core thesis — that agreement between 5-minute and 1-hour BTC momentum predicts KXBTC15M repricing within a $0.25–$0.75 band — is rejected by this round of testing. The strategy type does not earn its keep on this market at this timeframe.

Each variant remains saved in the Turbine strategy library for auditability, but I would not allocate risk capital to any of these configurations.

Long Disclaimer

This report is a historical simulation research document produced by Turbine's research tooling. It is intended for internal use only and does not constitute investment advice, a recommendation to trade, or a solicitation of any kind.

All performance figures are derived from backtesting against historical market data as of the date of the simulation. Backtested results do not represent actual trading and are subject to inherent limitations, including but not limited to: survivorship bias, look-ahead bias, fill assumptions, fee modeling, liquidity assumptions, and the inability to perfectly replicate real-world execution conditions.

The Kalshi market referenced (KXBTC15M) is a binary outcome contract on BTC price movement over 15-minute windows. Trading such instruments involves substantial risk, including the risk of total loss of principal. Crypto markets are volatile, and short-duration binary contracts amplify that volatility.

The parameter sweep and permutation test statistics included in this report are diagnostic tools. A permutation p-value below 0.05 does not necessarily indicate a tradable edge, particularly when accompanied by negative net PnL, negative Sharpe ratios, and fee drag exceeding gross profitability — all of which are present in this dataset. The permutation design used here permuted the edge feed timing but not the market price series, meaning price-dependent entry conditions were not fully stress-tested.

The deflated Sharpe ratio of 0.35 for the top variant is below the threshold that would distinguish it from the expected maximum of a skill-less parameter sweep. Accordingly, the top results are reported as consistent with selection noise and overfitting. The word "strong," "validated," or "promising" should not be applied to any variant in this report.

Finally, under no circumstances should past simulated performance be construed as indicative of future results. Markets change, regimes shift, and strategies that fail in backtests can perform differently in live conditions — usually not better, but differently. No capital should be allocated to any variant described herein without substantially more robust evidence.