Kalshi BTC 15m Extreme-Yes Strategy — Research Report

1. Short disclaimer

This report is a historical simulation study only. Nothing in this document is a promise of future profits, nor a recommendation to trade. All results are in-sample and subject to overfitting, fill-assumption risk, and selection bias. Read the long disclaimer at the end before making any decisions.

2. Intro / thesis

We tested the intuition that entering Kalshi’s BTC 15-minute binary markets near expiry, when the YES price is extremely high (≥ 0.80) and a directional BTC move confirms the view, can produce consistent profits when held to settlement. The baseline approach uses a 5-minute price change as the directional filter and requires the trade to be placed within the final five minutes before expiration. Our research explores whether a different entry window or a more reliable directional filter — EMA crossover, SMA crossover, VWAP, velocity, or raw momentum — would improve returns.

What we actually traded (derived from the strategy DSL):

• Asset: BTC
• Series ticker: KXBTC15M (Kalshi)
• Entry logic: Go long YES when contract price ≥ 0.80, time to expiry ≤ 5 minutes, and the 5-minute BTC price change (from Coinbase) is positive. Go long NO when contract price ≤ 0.20, time to expiry ≤ 5 minutes, and the 5-minute change is negative.
• Exit: Hold to expiry (positions are force-sold with 10 seconds remaining).
• Loop interval: 10 seconds.
• Risk controls: Variable price floor and ceiling across the sweep (see below), fixed 5-contract entry size.

3. Variant and strategy explanation

This was a 2‑D parameter sweep, not a filter-comparison study. We varied only two run-time risk parameters:

• risk.price_floor: tested at ten levels from 0.05 to 0.45.
• risk.price_ceiling: tested at ten levels from 0.55 to 0.95.

All 100 cells used the same underlying edge feed (Coinbase BTC-USD providing change_5m, EMA 12‑1m, SMA 20‑1m, price, VWAP 1h, and velocity 1m) and the same core rule structure. Different directional filters were not exercised — the sweep kept the 5‑minute momentum rule fixed. Every successful variant is saved as a runnable Turbine strategy.

The logic is straightforward: we only want to buy into an extreme price when the spot market is moving in the same direction, under the assumption that late, one‑sided moves are less likely to reverse before expiry.

4. Top results

The top eight variants are literally identical in performance because risk.price_floor had no effect in this historical window — not a single fill was blocked by a floor below 0.45. The winner’s coordinates are:

• risk.price_floor: 0.05 (but any floor ≤ 0.45 works identically here)
• risk.price_ceiling: 0.95

Saved strategies (top 8):

• btc-extreme-hold-4c9f792910e0 (floor 0.05 / ceil 0.95)
• btc-extreme-hold-784334cfb53c (floor 0.09 / ceil 0.95)
• btc-extreme-hold-b75f44d9c848 (floor 0.14 / ceil 0.95)
• btc-extreme-hold-95ffed3534f8 (floor 0.18 / ceil 0.95)
• btc-extreme-hold-05a45a9ccf6a (floor 0.23 / ceil 0.95)
• btc-extreme-hold-74aa7ea14dd4 (floor 0.27 / ceil 0.95)
• btc-extreme-hold-fcdec33a5739 (floor 0.32 / ceil 0.95)
• btc-extreme-hold-64633bf5a59d (floor 0.36 / ceil 0.95)

Important caveats (from our robustness checks):

• 52% of fills in the top cell were at prices < 0.10 or > 0.90. Those are exactly the price levels where executable fill assumptions are shakiest — in live trading, you may not get filled at all, or the spread may eat a large fraction of the notional edge.
• The deflated Sharpe of 0.61 is not distinguishable from the luckiest noise‑only trial in the sweep. This is a strong signal that the top result is consistent with selection noise.
• The permutation test (edge‑feed scramble) gave a p‑value of 0.003 — meaning the timing of the Coinbase feed relative to the Kalshi series was better than 99.7% of randomised versions. However, the market price series was not permuted, so the test only covers feed timing, not the price‑based entry condition.
• Performance increases monotonically with the price ceiling, making this an obvious corner solution — you’re essentially allowing all extreme‑yes entries and hoping price‑floor restrictions filter losers. In this window, they didn’t.

5. Bottom results

Every variant with a price ceiling ≤ 0.68 took zero trades, regardless of the price floor. That’s 36 of the 100 cells. Their statistics are uniformly flat:

The immediate takeaway: in this historical window, limiting entries to a ceiling ≤ 0.68 completely crippled the strategy. There simply were no contracts meeting the extreme‑price criterion that also fell below those tighter ceilings. This doesn’t mean the rule is “bad” permanently — it just means it removes all signal in this sample, leaving cash earning nothing.

6. Conclusion

We set out to see if a different entry window or a more reliable directional filter could improve on the baseline 5‑minute momentum approach. In practice, we tested only price‑floor and price‑ceiling variations — the directional filter was never altered. The historical results are dominated by a single parameter axis (ceiling) and an inert second axis (floor). The best‑performing combination is the least restrictive on the ceiling (0.95) and essentially any floor below 0.45.

The raw numbers look striking — an 884% ROI and a 93.6% win rate — but the robustness layer tells a much more cautious story:

• The deflated Sharpe of 0.61 is not strong evidence of genuine edge.
• Over half the fills sit at price extremes where execution risk is highest.
• The winner is a corner-solution: just raise the ceiling to let all trades through.

We must state plainly: the top results are consistent with selection noise and overfitting. No variant should be described as strong, validated, or promising based on this sweep alone. This is a useful starting point for out‑of‑sample testing, not a strategy ready for deployment.

7. Long disclaimer

This is a historical simulation study, not a live track record.

• All performance metrics are computed on past market data. They assume perfect fills with no slippage, no exchange downtime, no API latency, and no capacity constraints.
• The fill assumptions are least trustworthy at extreme prices (< 0.10 or > 0.90), where 52% of the top strategy’s fills occur. In live trading, those fills may be unavailable or far worse than the modelled price.
• We conducted a parameter sweep across 100 cells, which introduces severe selection bias. The best cell’s performance is almost certainly upward‑biased relative to future performance.
• The deflated Sharpe ratio (0.61) is below our internal threshold of 0.95, meaning we cannot reject the hypothesis that the top result arose from noise.
• The permutation test p‑value (0.003) measures only feed‑timing robustness, not the reliability of the price‑based rule itself.
• Trading binary options on prediction markets involves significant risk, including total loss of capital. Past simulated performance does not guarantee future results.

You should assume that any strategy described here will perform worse in live conditions — possibly substantially worse. This report is provided for informational purposes only and does not constitute investment advice. Consult your own research and risk tolerance before trading.

All successful variants are automatically saved as runnable Turbine strategies for further testing.