Research Report: Kalshi BTC 15-Minute Binaries — Late-Discount Momentum Strategy

Short Disclaimer

This is a historical simulation research note only. It does not constitute trading advice, and past simulated results do not imply future profits.

Intro / Thesis

The original thesis was that in the final 30–75 seconds of BTC 15-minute binary contracts on Kalshi, heavily discounted contracts (price under roughly 0.30) reprice sharply toward settlement when the 5-minute Coinbase trend and 1-minute velocity point in the contract's favor. The idea was to buy YES on rising momentum into the close and NO on falling momentum, take profit at a tier, and flatten before the final 30 seconds.

In practice, the strategy family that was swept was narrower: a custom KXBTC15M strategy using a 10-second loop, Coinbase BTC-USD price feed, price floors and ceilings, spread filters, late-entry thresholds, and stop-loss/take-profit rules. The simulation completed all 100 parameter cells.

The headline result is not strong. The entire family produced only one profitable cell. Most cells produced zero trades and zero PnL. The winning cell had high raw ROI but weak breadth, a low win rate, extreme fill assumptions, and a high likelihood of being selection noise.

Variant and Strategy Explanation

The base strategy is a custom Kalshi strategy on the KXBTC15M series. It uses:

• Asset: BTC
• Series ticker: KXBTC15M
• Loop interval: 10 seconds
• Market data: Coinbase BTC-USD, refreshed every 5 seconds
• Edge fields: price, 5-minute change, 1-minute velocity

Key base rules:

• Flatten all positions when time to expiry is 30 seconds or less.
• Tiered stop losses: sell all if unrealized PnL is below -$3.00 for positions of 5 or less, and below -$5.00 for positions of 6–10.
• Take profit: sell all if a YES position exists and the YES best bid is at or above 0.50.
• Buy YES late if price is below 0.30, spread is at or below 0.065, time to expiry is 60 seconds or less, 5-minute change is above 0.0010, and 1-minute velocity is above zero.
• Buy NO late under the same entry conditions, but with 5-minute change below -0.0010 and 1-minute velocity below zero.

The parameter sweep changed two risk controls:

• risk.price_floor: 0.05 to 0.45
• risk.price_ceiling: 0.55 to 0.95

That produced 100 completed cells. Each successful variant is saved as a runnable Turbine strategy.

Top Results

The top variant by net PnL was:

• Label: BTC strategy · floor 0.05 / ceil 0.55
• Strategy slug: deep-value-late-d3c8e8ec6811
• Total PnL: $12.56
• ROI: 125.6%
• Sharpe: 0.33
• Win rate: 15.79%
• Total trades: 38
• Max drawdown: -$0.18

This was the only variant in the sweep that generated any trades or PnL.

On the surface, the ROI looks attractive. But the trade-level picture is much weaker. The win rate was under 16%, meaning the positive PnL came from a small number of relatively large winners offsetting many losing trades. That is a fragile profile, not a broad edge.

The robustness statistics reinforce this concern. The deflated Sharpe for the winner is 0.9895, while the expected max Sharpe under the raw cell count assumption is 0.0831. That means the top result is entirely consistent with selection noise across the 100-cell sweep. The neighborhood degradation is 1, meaning the winner is isolated: immediately adjacent parameter cells collapse to zero trades and zero PnL.

There is also a fill-quality warning. The simulation flagged that 89% of winner fills were at prices below 0.10 or above 0.90, where fill assumptions are least trustworthy. The threshold for the warning was 30%. This does not invalidate the simulation, but it means the reported PnL likely overstates what would have been achievable in live markets.

I am not describing this variant as strong, validated, or promising. It is a single isolated positive cell in a sweep where 99 of 100 cells did nothing.

Bottom Results

The bottom results are not dramatic in dollar terms, but they are informative.

Most of the lower-ranked variants produced exactly zero trades:

• Floor 0.05 / ceil 0.59: 0 trades, 0 PnL
• Floor 0.05 / ceil 0.64: 0 trades, 0 PnL
• Floor 0.05 / ceil 0.68: 0 trades, 0 PnL
• Floor 0.05 / ceil 0.73: 0 trades, 0 PnL
• Floor 0.05 / ceil 0.77: 0 trades, 0 PnL
• Floor 0.05 / ceil 0.82: 0 trades, 0 PnL
• Floor 0.05 / ceil 0.86: 0 trades, 0 PnL
• Floor 0.05 / ceil 0.91: 0 trades, 0 PnL

The bottom of the sweep is essentially a flatline. The only reason the top variant ranks first is not that the lower variants lost money. It is that they never entered a trade.

This is a key structural finding. The entry conditions in the base strategy are narrow enough that most parameter combinations never trigger. The floor/ceiling combination of 0.05/0.55 was the only one that allowed any activity. Everything above that ceiling range was inert.

From a robustness perspective, this is not a sign of a stable edge. A stable strategy would typically show some activity, and often some losses, across neighboring parameter values. Here, the activity drops to zero the moment the ceiling moves from 0.55 to 0.59.

Conclusion

This sweep does not support the original thesis in any reliable way.

The base strategy was coherent: late-entry discounted binaries, momentum confirmation, tight spreads, early flattening, tiered risk exits. But the empirical results are thin. The only profitable variant had low win rate, extreme fill prices, and an isolated parameter location. The deflated Sharpe was far above the expected max Sharpe for a 100-cell sweep, which means the result is consistent with overfitting or selection noise.

The broader family was essentially inactive. 99 of 100 cells produced zero trades. The edge was not distributed across the parameter space. It appeared once, at one corner, under fill assumptions that are unlikely to hold in live trading.

The honest read is that this is not a validated strategy. It is an isolated historical artifact that deserves no forward-looking confidence. If this idea is revisited, it would need a different entry framework, better fill modeling, and evidence of activity across more than a single parameter cell.

Long Disclaimer

This report is provided for research and informational purposes only and does not constitute an offer, solicitation, recommendation, or endorsement of any security, financial product, instrument, or trading strategy.

The analysis presented here is based on historical simulation and backtesting. Backtested results are inherently limited. They do not represent actual trading and do not account for real-world frictions such as live liquidity constraints, partial fills, order queue position, exchange outages, latency, slippage, fee schedules beyond those modeled, or the possibility that quoted prices at the time of a simulated fill were not actually executable.

The strategy described uses late-entry binary contracts with short time horizons. Binary event contracts can experience rapid and discontinuous price moves, especially near expiration. Fill assumptions in such environments are highly sensitive to venue liquidity and market microstructure. The simulation itself flagged that the winning variant's fills were concentrated at extreme prices, where execution assumptions are least reliable.

The robustness statistics should be read carefully. The top result's deflated Sharpe was above the expected maximum Sharpe for the number of cells tested, indicating that the result is statistically consistent with selection noise. No variant in this report should be considered validated, robust, or predictive of future performance.

Kalshi and Coinbase are third-party platforms. This report is not affiliated with or endorsed by either. Market data may be delayed, corrected, or revised. Strategy parameters were fixed for research purposes and were not optimized on live data in this exercise.

Past performance, whether simulated or real, is not indicative of future results. Trading binary options and crypto-related instruments involves substantial risk, including the possible loss of principal. Anyone considering these strategies should consult a qualified financial professional and fully understand the risks involved.

Finally, this report may include forward-looking statements. Those statements are based on assumptions that may prove incorrect, and actual outcomes may differ materially. No representation or warranty is made as to the accuracy, completeness, or fitness for any purpose of the information contained herein.