Kalshi KXBTC15M Research Report

Short Disclaimer

This is historical simulation research only. Past simulated performance does not imply future results.

Intro / Thesis

We tested a directional edge on Kalshi's KXBTC15M market: using Coinbase BTC price relative to its 1-hour VWAP, combined with 5-minute momentum, to predict whether a 15-minute contract settles YES or NO. The base idea was that price trading above VWAP with rising 5-minute momentum favors buying YES, while price below VWAP with falling momentum favors buying NO.

Variant and Strategy Explanation

The strategy runs on a 30-second loop, checking Coinbase BTC-USD price, 1-hour VWAP, and 5-minute percentage change. Entry rules require:

• Buy YES when BTC price > 1h VWAP and 5m change > +0.125%
• Buy NO when BTC price < 1h VWAP and 5m change < -0.125%

Additional filters included a spread cap of $0.03, price band constraints, and exits near contract close or on stop-loss.

For this research sweep, we varied only the price floor and price ceiling across 100 combinations. Floor ranged from 0.05 to 0.45. Ceiling ranged from 0.55 to 0.95. Every successful variant from this sweep is saved as a runnable Turbine strategy.

Top Results

The best-performing variant used a floor of 0.05 and ceiling of 0.55, producing:

• Net PnL: $32.39
• ROI: 64.78%
• Max drawdown: -$4.66
• Sharpe: 0.63
• Trades: 699
• Win rate: 47.6%

The runner-up, with floor 0.45 and ceiling 0.73, generated $6.43 net PnL and 12.86% ROI, with a much lower Sharpe of 0.11. The next several variants cluster in the $4–6 net PnL range, all with Sharpe ratios below 0.15.

The standout result comes entirely from the most restrictive ceiling tested (0.55). No other ceiling produced net PnL above $6.43, and most positive results required low floors paired with tighter ceilings. This concentration of performance in a single corner of the parameter grid matters for interpretation.

Bottom Results

The worst variants consistently used the widest ceiling (0.95) across various floors. The bottom performer, with floor 0.18 and ceiling 0.95, lost $13.93 with a -27.86% ROI and max drawdown of -$23.37. Several nearly identical variants using ceiling 0.95 produced losses in the -$12 to -$14 range.

Notably, many losing variants had higher win rates than the top performer. The #100 variant won 54.6% of trades yet lost money, while the #1 variant won only 47.6% and earned $32.39. This is the classic signature of adverse selection: wide price bands allowed entries into contracts where the payout asymmetry did not compensate for the higher loss severity on losing trades.

Conclusion

The raw spread between best and worst variants is large: $32.39 profit versus -$13.93 loss. However, the robustness statistics tell a more cautious story.

Deflated Sharpe: 0.99
Expected max Sharpe under selection noise: 0.30
Neighborhood degradation: 1.04

The deflated Sharpe exceeds the expected max Sharpe, which ordinarily would indicate some signal survives the parameter sweep. However, the neighborhood degradation of 1.04 is cause for concern: it means the top variant's edge deteriorates very quickly when you move to adjacent parameter cells. In plain terms, the best result is an isolated spike — its edge does not extend to nearby floor/ceiling combinations.

The marginal means confirm this. Average performance across most floor and ceiling values is flat-to-negative. Only the tightest ceiling (0.55) shows meaningful positive mean, and that is propped up almost entirely by the single outlier variant at floor 0.05.

I cannot describe the top variant as strong, validated, or promising. The performance concentration, the rapid degradation across neighboring parameters, and the flat-to-negative marginal returns across most of the grid are consistent with a parameter fit that happens to land on a favorable pocket of historical data. The strategy family as a whole does not demonstrate a reliable directional edge on KXBTC15M.

That said, the full sweep is saved in Turbine. The specific runnable strategy IDs are preserved for all 100 variants, allowing for continued monitoring if market conditions shift.

Long Disclaimer

This report is a historical simulation study produced for internal research purposes. It is not investment advice, a recommendation, or a solicitation to trade any security, contract, or prediction market instrument.

The results presented here are derived from backtesting against historical data. Backtested performance does not represent actual trading, does not account for all real-world frictions (such as liquidity constraints, partial fills, execution delays, or fee structures beyond those modeled), and may be materially different from live trading results.

The parameter sweep methodology used in this study involves testing many variations of the same underlying strategy. This inherently increases the probability of finding seemingly favorable results due to random chance, a phenomenon known as multiple testing or selection bias. The deflated Sharpe ratio and neighborhood degradation statistics are reported to help assess this risk, but they do not eliminate it.

A permutation test was not available for this run, so no permutation p-value is reported. The absence of a permutation test means we cannot quantify the probability that the observed top results arose purely from chance across the parameter grid.

All figures, including PnL, ROI, Sharpe ratios, win rates, and drawdowns, are simulated outputs from the Turbine research environment. They should not be interpreted as predictive of future performance. Markets change, edges decay, and parameters that worked historically may fail in live conditions.

Any strategy ID or strategy slug mentioned in this report refers to a saved historical simulation configuration within Turbine. These are research artifacts, not live deployments.