Kalshi KXBTC15M Velocity Strategy Research Report

Short Disclaimer

Historical simulation only. No live trading occurred. Results below may be unreliable due to extremely low trade counts. This is not investment advice.

Intro / Thesis

This research tested whether short-dated Bitcoin 15-minute binary markets on Kalshi could be traded using Coinbase 1-minute velocity as an entry signal. The core idea: when BTC velocity crosses a threshold in either direction, buy the side of the market that aligns with that velocity, but only if the price is between 50c and 90c. Each market gets one entry maximum, and positions are held to settlement. The sweep varied the price floor and ceiling to see which bounds produced the best fill rate, win rate, and settlement-direction accuracy.

The honest summary up front: the dataset produced almost no trades. Most variants never executed a single position. The few that did took one trade and lost. There is no meaningful performance to report.

Variant and Strategy Explanation

The base strategy observed KXBTC15M markets on a 10-second loop. It pulled Coinbase BTC-USD 1-minute velocity every 5 seconds. The rules were:

• Cancel all orders if time to expiry is 1 minute or less.
• Buy YES if velocity is above +0.50 and the YES best bid is between the price floor and ceiling.
• Buy NO if velocity is below -0.50 and the NO best bid is between the price floor and ceiling.
• Hold to settlement otherwise.

The sweep varied two parameters:

• risk.price_floor: 0.05 to 0.45 in 10 steps
• risk.price_ceiling: 0.55 to 0.95 in 10 steps

This created 100 variants. All 100 completed successfully.

A critical note on the parameter sweep: the base rules in the DSL referenced yes_best_bid and no_best_bid with floors and ceilings around 0.50–0.90. But the sweep widened the floor down to 0.05 and pushed ceilings from 0.55 to 0.95. The result was that very few price observations ever landed inside these widened bounds with a simultaneous velocity trigger. The market simply didn't offer qualifying entries during the simulated window.

Top Results

The top 8 variants are all effectively identical in outcome: either zero trades or one losing trade.

The "top" variants rank first only because they did nothing. A zero-trade strategy has zero drawdown and zero losses, which technically ranks above a one-trade strategy that lost. That is not a signal of quality.

The single trade taken by variants 5–8 was a loss. It produced a small drawdown of 4.79 (presumably in cents or normalized units), a Sharpe of 0.43 based on that single observation, and a win rate of 0%. One trade is statistically meaningless.

Every successful variant is saved as a runnable Turbine strategy, but saving a strategy that never trades is not the same as finding an edge.

Bottom Results

The bottom results are the same as the top results, because the entire sweep produced essentially no differentiated outcomes. There is no meaningful spread between the best and worst variants. The full list of bottom variants mirrors the top list: four variants with zero trades, four variants with one losing trade.

The robustness statistics confirm this flatness. The parameter sweep winner was floor 0.05 / ceiling 0.55 with a net PnL of 0. Every cell in the grid produced either 0 or an identical losing trade. The marginal means across all floor values and all ceiling values were flat at 0.

Three warnings are worth stating plainly:

1. Trade floor warning: The winner had 0 resolved trades. The reliability threshold is 30. This is far below any standard for statistical confidence.
2. PnL days floor warning: The winner had 1 distinct PnL day, below the threshold of 10. Daily Sharpe cannot be trusted on one day.
3. Inert axis warning: Varying risk.price_floor produced identical trades in every cell. The parameter was never actually exercised. The flatness here is not robustness—it's the absence of any signal at all.

The permutation test returned a p-value of 1.000. This means the strategy's edge-feed timing beat 0.0% of 1,000 time-scrambled re-sweeps. In plain terms: the results are entirely consistent with selection noise. There is no evidence that the velocity signal had any predictive value during this window.

Conclusion

This sweep did not find a tradable edge. The core problem is simple: the strategy almost never entered a position. Out of 100 variants, 96 never traded. The 4 that did each took exactly one trade, and that trade lost.

The thesis—that Coinbase 1-minute velocity crossing a threshold in combination with a 50c–90c price band would produce profitable entries—was not supported. The price and velocity conditions rarely aligned during the simulated period. Whether that's because the market was too quiet, the thresholds were not hit, or the price bounds were too broad is unclear from the data. What is clear is that the strategy produced no useful information.

The p-value of 1.000 from the permutation test is a direct statement that these results are consistent with noise. No variant should be described as promising, validated, or strong. The only honest conclusion is that this particular configuration, in this particular window, did not trade enough to evaluate.

A future iteration would need to address the fill rate problem first. Without a meaningful number of entries, no downstream metric—win rate, settlement accuracy, PnL, Sharpe—can be trusted.

Long Disclaimer

This report is a historical simulation research artifact produced by Turbine's automated research pipeline. It does not constitute investment advice, a recommendation to buy or sell any financial instrument, or a guarantee of future performance. All performance metrics are derived from historical data and may not reflect what would have occurred in live trading.

Key limitations of this research:

• Zero or minimal trades: The majority of variants executed zero trades. Any metric derived from zero or one trade is statistically meaningless and should not be interpreted as evidence of strategy quality.
• No live trading: No real capital was deployed. Simulated fills, fees, slippage, and liquidity may differ materially from live conditions.
• Permutation test scope: The permutation test scrambled the edge-feed timing but did NOT permute the market price series. The p-value of 1.000 means the velocity signal showed no predictive value relative to time-scrambled versions of itself, but it does not test price-based conditions in the rule.
• Inert parameters: The risk.price_floor axis was never effectively exercised. The sweep's flatness reflects a lack of entry opportunities, not robustness of the strategy.
• Short observation window: The strategy had 1 distinct PnL day, well below any reasonable threshold for assessing daily Sharpe or drawdown characteristics.
• Overfitting risk: With 100 variants and virtually no differentiating trades, any apparent top performer is likely an artifact of sorting noise, not a real edge.

Historical simulation software has inherent limitations. Past results, including simulated past results, do not guarantee future outcomes. Before making any trading decision, you should conduct your own research and consult with a qualified financial professional. Turbine makes no representation or warranty, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of this research.