Research Report: Kalshi KXBTC15M Multi-Horizon BTC Trend Confirmation

Short Disclaimer

This is historical simulation research only. Nothing here implies future profits or predictive ability. Past backtest results are not indicative of live trading outcomes.

Intro / Thesis

This report covers the fourth validation pass on a Kalshi crypto strategy trading the KXBTC15M series. The thesis was simple: when BTC's 5-minute and 1-hour returns agree in direction, Kalshi's 15-minute BTC markets may be mispriced relative to that momentum. Buy YES when both timeframes are positive. Buy NO when both are negative. Filter for reasonable entry prices and tight spreads, and exit before the final two minutes.

The idea was to test whether multi-horizon trend confirmation produces repeatable historical edge. Specifically, we wanted to know if agreement between short-term and medium-term momentum predicts near-term repricing in the binary market.

We ran 100 completed variants. Every single one lost money.

Variant and Strategy Explanation

The base strategy is a custom Kalshi loop running every 30 seconds on the most liquid KXBTC15M contract. It pulls BTC spot data from Coinbase with a 10-second refresh, monitoring both 5-minute and 1-hour percentage changes.

The entry logic:

• Buy YES when BTC change over 5m AND 1h are both positive.
• Buy NO when BTC change over 5m AND 1h are both negative.

Entries are constrained to:

• Price between $0.35 and $0.65 (with some variants widening the band).
• Spread of $0.03 or less (some variants tightened to $0.01).
• More than 2 minutes remaining until expiry.

The strategy also has two flatten rules: exit everything if unrealized PnL hits -$25, or if fewer than 2 minutes remain.

The completed sweep varied two parameters:

• entry_low: $0.25, $0.30, $0.35, $0.40, $0.45
• entry_high: $0.55, $0.60, $0.65, $0.70, $0.75

That yields 25 unique parameter combinations. Each also ran against several spread caps, producing 100 total variants. Each successful variant is saved as a runnable Turbine strategy with its own strategy ID and slug.

Top Results

The best-performing variant, ranked #1 with a total PnL of -$4.02, was:

• Label: BTC strategy · band 0.25–0.55 · spread cap 0.02
• ROI: -4.02%
• Sharpe: -0.06
• Max drawdown: -29.75
• Total trades: 4,165
• Win rate: 51.1%

That win rate sounds okay at first glance. But the PnL is negative even with slightly more winners than losers. The typical winning trade was smaller than the typical losing trade, and round-trip costs ate the difference.

More importantly, the top eight variants all produced nearly identical results — the same -4.02 PnL, same drawdown, same win rate. That flatness is the key finding. Changing the entry band or spread cap did almost nothing in this window. The warning notes are explicit about this: both entry_low and entry_high appear to be inert axes. Varying them produced identical trades, meaning the parameter sweep never actually exercised different market conditions. The observed flatness is not evidence of robustness; it's evidence that the rules triggered the same way regardless of band settings.

The deflated Sharpe for the best variant is 0.37, well below the 0.95 threshold for significance against the luckiest skill-less sweep. The best result here is consistent with selection noise, not repeatable edge.

Bottom Results

The worst variant was:

• Label: BTC strategy · band 0.25–0.55 · spread cap 0.01
• ROI: -7.57%
• Sharpe: -0.06
• Max drawdown: -44.18
• Total trades: 3,836
• Win rate: 46.7%

The tighter spread cap reduces trade count but doesn't improve quality. Win rate drops to 46.7%, and the drawdown deepens to -44.18. The tighter spread filter removes some trades, but those removed trades were apparently not the main source of loss — the remaining trades just lose faster.

Across the entire sweep, the bottom variants confirm the same conclusion: this edge, as specified, does not exist in the historical window.

Robustness Assessment

The permutation test ran 100 re-sweeps against time-scrambled versions of the edge feed. The headline result is p = 0.010, meaning the real feed beat 99% of the scrambled versions. That sounds meaningful at first glance. But the test design has an important caveat: it permutes the edge feed timing but anchors trade prices, so price-based filters are not truly tested by the permutation. The p-value tells us the timing of BTC trend signals had some relationship to outcomes, but it does not rescue the headline result.

The p-value of 0.010 is below the 0.05 threshold, but the deflated Sharpe of 0.37 is far below 0.95. When deflated Sharpe is near expected_max_sharpe, the correct interpretation is that the top result looks like the best draw from a noisy sweep, not a validated strategy.

The robustness warnings flag three issues clearly:

1. Fee drag: Fees consume 105% of the winner's gross PnL — the entire gross edge, had there been one, was more than eaten by costs.
2. Low deflated Sharpe: 0.37, essentially noise-level.
3. Inert axes: Both swept parameters produced identical trades, so the sweep offers no evidence of parameter sensitivity or robustness.

The statistics rules for this report apply plainly: the top results are consistent with selection noise and overfitting. I won't describe any variant as strong, validated, or promising.

Conclusion

This was a negative result, and it deserves to be reported that way. Multi-horizon BTC trend agreement — at least as implemented here with these entry bands, spread caps, and time-to-expiry filters — does not produce a positive historical PnL in Kalshi KXBTC15M markets.

The best variant lost 4.02%, and the best variant's performance is not distinguishable from noise after accounting for the number of parameter combinations tried. The sweep itself didn't exercise its parameters meaningfully, so there's no evidence that a better band or spread setting exists in this family. The strategy generates lots of trades (4,165 in the top case) with a coin-flip win rate, and the economics are negative after fees.

We can say the following with confidence:

• The base thesis, as tested, fails.
• Tighter spreads reduce trades but don't improve quality.
• The entry bands in this sweep were functionally irrelevant.
• Max drawdowns of -29.75 to -44.18 are too deep for a strategy with no gross edge.

If this family is revisited, the next step would be to rework the signal itself — perhaps different lookback windows, a different confirmation mechanism, or an entirely different price-structure trigger. What won't help is another sweep of entry bands in the same range.

Long Disclaimer

This document is a historical simulation research report prepared by Turbine's research function. It is provided for informational and analytical purposes only and does not constitute investment advice, a recommendation to buy or sell any security or derivative, or an offer to provide investment management services.

All performance figures are derived from backtesting on historical data. Backtested results are inherently limited and do not represent actual trading. They do not account for every real-world friction, including but not limited to latency, partial fills, exchange outages, market impact, and changes in market microstructure. Backtests can be overfit, and results that appear stable in-sample often fail out-of-sample.

The strategy described here is a custom automated trading system for Kalshi's KXBTC15M market. Kalshi markets are event contracts and carry their own regulatory and platform-specific risks. This report does not assess the suitability of Kalshi or any other venue for any user.

The permutation test referenced in the robustness section permutes edge-feed timing but does not permute market price series. Its p-value cannot be interpreted as a full test of the price-based entry conditions. The deflated Sharpe statistic referenced herein is below the threshold at which results would be considered distinguishable from selection noise, and the report explicitly concludes that these results are consistent with overfitting and are not promising.

Every successful variant in the sweep has been saved as a runnable Turbine strategy. This does not imply that any of those strategies should be deployed with real capital. Deploying any automated strategy involves risk of loss, and past performance — including historical simulation performance — is not indicative of future results.

No future profits are implied anywhere in this report. Any references to edge, signal, or predictability are qualified by the negative results presented and the robustness warnings issued. Readers should consult their own advisors and conduct their own diligence before engaging in any trading activity.