Research Report: Coinbase Momentum v2 — KXBTC15M Parameter Sweep

Venue: Kalshi
Market: KXBTC15M (BTC 15-minute binary)
Date: April 2025

1. Short disclaimer

This is a historical simulation research report. Nothing in this document constitutes trading advice or a prediction of future returns. Past performance, especially from parameter sweeps, routinely overstates what any single configuration would have achieved live.

2. Intro / thesis

We set out to test whether stronger, carefully timed Coinbase momentum signals could produce reliable entries on the KXBTC15M market. The base strategy uses 5-minute BTC change and 1-minute velocity thresholds to buy YES or NO contracts during a narrow window of remaining contract time. The core idea: by only acting when Coinbase shows clear directional push and the contract is neither too early nor too close to settlement, we avoid low-conviction noise.

The parameter sweep varied the price floor and ceiling that gate entries, while keeping the entry timing, spread limits, and exit rules fixed across all runs. One hundred variants were explored, ranging from wide price bands to very tight ones, to understand which combinations held up best historically.

3. Variant and strategy explanation

All variants share the same base logic:

• Asset traded: BTC binary (KXBTC15M)
• Loop interval: 10 seconds
• Data source: Coinbase BTC-USD (price, 5-minute change, 1-minute velocity), refreshed every 5 seconds
• Entry conditions:
  • YES entry: change_5m > 0.001 AND velocity_1m > 0
  • NO entry: change_5m < -0.001 AND velocity_1m < 0
  • Price must fall within [price_floor, price_ceiling]
  • Spread ≤ 1.5¢
  • Time remaining between 4 and 10 minutes
  • No existing position
• Position size: 25 contracts per entry
• Exits: Sell all on unrealized PnL ≤ -$25, unrealized PnL ≥ $15, or when time remaining drops to 2 minutes

The only parameters varied across the sweep were the price floor (0.05 to 0.45 in ten steps) and price ceiling (0.55 to 0.95 in ten steps), creating a 10×10 grid of 100 configurations. Every successful variant is saved as a runnable Turbine strategy.

4. Top results

The sweep returned an unusually flat top tier. Eight variants tied at the maximum net PnL of $93.79, all sharing the same price ceiling of 0.73 but with floors ranging from 0.05 up to 0.36. The best performer by rank:

Importantly, variants from floor 0.05 through 0.36 all produced identical results at the 0.73 ceiling. This clustering — eight identical outcomes across a wide range of floor values — suggests the floor parameter is not driving differentiation within this band; the ceiling is the binding constraint for these winners.

However, the robustness statistics raise serious concerns:

• Deflated Sharpe equals 0.87, which falls below the 0.95 threshold we use to distinguish signal from selection noise. The expected maximum Sharpe from a pure noise sweep of this size is 0.51, and the deflated value is only modestly above that.
• P-value from the permutation test is 0.034. This means the strategy's edge-feed timing beat 96.6% of time-scrambled versions of itself, which is statistically significant but far from overwhelming given the number of variants tested.
• Only 7 distinct PnL days support the winner's Sharpe, well below the 10-day minimum we prefer for daily-level confidence.

The honest read: the top results look good in raw numbers but are consistent with what you would expect from a moderately sized parameter sweep over a favorable historical window. We cannot describe any variant as validated or robust based on this data alone.

5. Bottom results

The worst performer landed at -$9.15 net PnL with a -18.3% ROI, using floor 0.05 and ceiling 0.55. This variant took only 2 trades and lost both. The extremely tight upper price bound effectively prevented the strategy from participating in most momentum moves, starving it of entry opportunities.

The bottom tier more broadly shows a pattern: variants with very high ceilings (0.91 and 0.95) combined with low or mid-range floors tended to produce larger drawdowns — one cluster saw max drawdowns of -$43.12 — even when total PnL remained positive. High ceilings let the strategy enter at elevated prices where mean reversion and late-contract noise exact a heavier toll. Win rates in the bottom ranks were only marginally worse than the winners, but the damage per losing trade was larger.

Notable bottom performers:

The 0.91–0.95 ceiling cluster produced positive PnL but with deeply uncomfortable drawdowns relative to return.

6. Conclusion

The Coinbase Momentum v2 strategy on KXBTC15M shows attractive headline numbers in its best historical configuration — 187% ROI, 0.92 Sharpe, 63% win rate — but the robustness checks deflate that enthusiasm considerably. The top eight variants are functionally identical outliers tied to a single ceiling value, the deflated Sharpe sits below our confidence threshold, and the permutation p-value of 0.034, while technically significant, does not clear the high bar needed after testing 100 configurations.

Price ceiling emerges as the dominant driver of outcomes. A cap near 0.73 kept the strategy in a sweet spot where momentum entries occurred at prices that still had room to run. Ceilings above roughly 0.77 degraded net PnL and increased drawdowns. The price floor, by contrast, had almost no discriminatory power across a broad range when the ceiling was held at 0.73.

We would not deploy any single variant from this sweep with high conviction. The natural next step would be out-of-sample testing — ideally on a held-out time period not touched during this sweep — to see whether the 0.73 ceiling cluster maintains any edge or collapses toward the mean.

7. Long disclaimer

This report presents the results of a historical simulation conducted for research purposes on Turbine's platform. No live trading occurred. All performance figures are hypothetical and subject to known limitations of backtesting, including but not limited to survivorship bias, look-ahead bias where data handling permits, and overfitting to the specific historical period studied.

The parameter sweep tested 100 variants. When many configurations are tested, the best observed results are almost certainly upward-biased estimates of true out-of-sample performance. The deflated Sharpe ratio and permutation test are attempts to quantify this bias, but they do not eliminate it. A deflated Sharpe below 0.95 indicates that the top performer's edge, after adjusting for the multiplicity of tests, does not reach conventional thresholds for statistical reliability.

The permutation test shuffled only the edge-feed values — Coinbase BTC change and velocity — while leaving market prices and timestamp-cued conditions intact. This isolates whether the timing relationship between momentum signals and trade entries adds value beyond random alignment. The test does not assess whether the market price series itself contains exploitable structure, which is a separate question requiring other methods.

Price-based entry conditions (floor and ceiling) were held fixed per variant during the permutation, so their contribution to performance was not directly stress-tested by that procedure. Readers should interpret the marginal analysis and neighborhood degradation metric as complementary evidence.

This report does not recommend, endorse, or predict the future performance of any strategy, parameter set, or market. Trading binary options involves the risk of total loss. Consult your own judgment and risk tolerance before allocating capital. Past simulation results are not indicative of future results.