Daily BTC Market-Making Strategy on Kalshi: Parameter Robustness Study

Short Disclaimer

This report presents historical simulation results only. Past simulated performance does not guarantee future outcomes. All strategies carry risk of loss, and the findings here may reflect selection noise rather than genuine edge.

Intro / Thesis

Daily Bitcoin markets on Kalshi occasionally present spreads wide enough that a trader can quote both YES and NO orders simultaneously at the best bid and, if filled, capture a positive expected spread without needing to predict which way BTC moves. This study tests whether a simple post-only maker strategy — one that quotes both sides while flat, stops quoting once filled, holds to settlement, and uses a mechanical $25 stop-loss — would have shown profitable results historically.

We ran a full grid search across 100 parameter combinations, varying the price floor (minimum contract price at which the strategy participates) and price ceiling (maximum participation price). Every combination completed its backtest. The headline results show the top-ranked variant achieving an 85.5% return on initial capital, but the robustness statistics tell a more cautious story.

Variant and Strategy Explanation

The base strategy is straightforward. It trades the KXBTCD daily BTC markets on Kalshi. Every 60 seconds, it checks three conditions:

• The account must be flat (no open position).
• The bid-ask spread must be at least $0.02 wide.
• The contract must have more than one minute remaining until expiration.

When all conditions are met, it places two post-only limit orders: one buying YES at the best bid, one buying NO at the best bid. Both are $25 in size. Once either side fills, quoting stops. The position is held to expiration — unless unrealized losses hit $25, at which point the entire position is sold immediately.

The variants explore different boundaries for trade entry. The price_floor parameter (tested from $0.05 to $0.45) sets a minimum contract price below which the strategy won't quote. The price_ceiling (tested from $0.55 to $0.95) sets the upper limit. The idea is to filter out contracts trading near the extremes, where one side is near $0.01 or $0.99 and the risk of a sudden move through the stop-loss is higher, versus contracts in the middle of the range where both sides have meaningful value.

The grid tested every combination of those ten floor values and ten ceiling values, for 100 total variants. Each completed backtest represents a specific saved Turbine strategy that could theoretically be deployed.

Top Results

The best-performing combinations cluster around lower price ceilings ($0.55) paired with several different floor values. Here are the top eight variants:

The rank 1 and 2 variants are identical in outcome because they only differ in the floor value (0.05 vs 0.09), and that difference was never tested in practice — no contracts traded at prices between those two thresholds that would have changed the strategy's behavior. The same dynamic appears in other matched pairs throughout the results.

The standout observation is the top variant's shallow drawdown relative to its return. At -$7.70 max drawdown on $42.76 total profit, the worst-to-best ratio looks attractive on the surface. Win rates hover around 43-60%, meaning the strategy loses money on a significant fraction of individual trades but the positive outcomes are large enough to carry the total P&L.

Bottom Results

The worst performers share a common feature: high price floors combined with relatively tight ceilings. Here are the bottom eight:

The worst variant (floor 0.23, ceiling 0.68) lost over half of its allocated capital across just 10 trades. The common thread in these losers is that the tighter participation window reduced the number of opportunities and concentrated risk into fewer contracts at prices where the spread-capture edge may have been insufficient to overcome adverse selection.

Notice that the rank 93 variant (floor 0.45, ceiling 0.59) took only 3 trades and won none of them — zero percent win rate across an extremely narrow slice of the market. This is a stark illustration of what happens when the parameter filter is too restrictive: the strategy sits out most opportunities, then catches only the worst ones.

Conclusion

The raw P&L curve looks appealing in the best case, but the robustness statistics strongly caution against reading these results as evidence of a real edge. Two specific warnings from the analysis deserve attention.

First, the top-ranked variant produced only 16 resolved trades. That is a small sample. With fewer than 30 trades, standard performance metrics like Sharpe ratio and win rate are statistically unreliable. A few chance outcomes can dominate the entire record.

Second, and more importantly, the deflated Sharpe ratio — a metric that adjusts for the fact that we tested 100 different parameter combinations — comes in at 0.44. That number sits below 0.95, meaning the best result we found is not statistically distinguishable from what we would expect to see by pure luck when running this many tests. The marginal analysis reinforces this: while lower ceilings and lower floors tended to produce better average outcomes, the neighborhood degradation score of 0.17 indicates that small changes in parameters produce meaningful differences in results, which is itself a sign that the surface is noisy rather than structured.

The permutation test was not available in this run, so we cannot quote a specific p-value, but the deflated Sharpe alone is sufficient to say that these results are consistent with selection noise.

All that said, the economic logic of the strategy is sound in principle — if spreads are wide enough and the market maker can avoid being picked off on adverse moves, capturing spread is a viable business. The simulation simply does not provide enough evidence to assert that this particular implementation would deliver those results going forward. Anyone considering deploying one of these saved Turbine strategies should treat the historical numbers as a sketch, not a forecast, and should expect live performance to differ meaningfully.

Long Disclaimer

This research report is produced for informational and educational purposes only. It describes historical simulations run on completed market data and does not constitute investment advice, a trading recommendation, or an offer to buy or sell any financial product.

No forward-looking claims. All performance figures — including total P&L, ROI, Sharpe ratios, win rates, and drawdown statistics — are backward-looking results from parameter sweeps. Past simulated performance does not predict future returns. Live trading involves execution risk, latency, market impact, regulatory considerations, and other factors not captured in historical backtests.

Survivorship and selection bias. The parameter sweep methodology tests many combinations and selects the best retrospectively. The deflated Sharpe ratio reported here attempts to quantify the magnitude of this selection bias, but no statistical adjustment can fully compensate for the fact that the worst-performing combinations are ignored and only the best are highlighted. Results from any single variant are likely to overstate expected future performance.

Small sample limitations. Several top-ranked variants executed fewer than 30 trades. Performance statistics derived from small samples have wide confidence intervals and limited predictive content. A