ETH 15-Minute Triple-Confirmation: Does the Tradable Price Band Matter?

Turbine research note — Kalshi, crypto (ETH), series KXETH15M

Disclaimer (short)

This is a historical simulation study, not investment advice. Past simulated results do not predict future outcomes, and nothing here should be read as a promise of profit. Please read the full disclaimer at the bottom before drawing any conclusions.

Intro / thesis

The question behind this sweep is narrow and practical: on Kalshi's 15-minute ETH contract market (series KXETH15M), does the price band you let the strategy trade in — the risk.price_floor and risk.price_ceiling — actually change the results of a specific entry pattern? Or is the entry filter doing all the work, with the band just along for the ride?

The entry pattern in question is a triple-confirmation setup. The strategy checks every 10 seconds. It will only open a position when it is flat, the contract has 10 minutes or less to expiry, and three separate crypto feeds are moving the same direction: ETH, BTC, and SOL all showing a 5-minute change beyond ±0.1%. On top of that, ETH must be trading above its 1-hour VWAP and its 12-period EMA (for a YES entry) or below both (for a NO entry). The contract itself has to be priced between 0.35 and 0.65 with a spread no wider than 0.015, and the entry size is 50 contracts. Exits are tiered take-profits scaled by position size (small, medium, large), plus a stop that triggers below -5 unrealized PnL when the contract trades between 0.15 and 0.85. The sweep varied only the floor and ceiling across ten values each, giving a 10×10 grid of 100 completed variants. The thesis: the band is a meaningful lever, not a cosmetic one.

Variant and strategy explanation

All 100 variants share the same rules; only the outer price band changes. Here is what actually runs each tick, in order:

1. Take-profit, small position. If size is ≤10, unrealized PnL is >3, and the contract trades between 0.10 and 0.90, sell everything.
2. Take-profit, medium position. If size is 11–20, unrealized PnL is >5, and the contract trades between 0.10 and 0.90, sell everything.
3. Take-profit, large position. If size is ≥21 and unrealized PnL is >8, sell everything.
4. Stop loss. If unrealized PnL is below -5 and the contract trades between 0.15 and 0.85, sell everything.
5. YES entry. When flat, with ≤10 minutes to expiry, ETH/BTC/SOL 5-minute changes all above +0.1%, ETH price above its 1-hour VWAP and 12-period EMA, contract price 0.35–0.65, and spread ≤0.015, buy 50 YES.
6. NO entry. The mirror image: all three feeds' 5-minute changes below -0.1%, ETH below VWAP and EMA, same price and spread window, buy 50 NO.

The position limit is 60 contracts, and the tradable band under test sits between a 0.01 floor / 0.99 ceiling at the wide end and narrower bands like 0.44 / 0.56 at the tight end. The window is 30 days of history. Each successful variant is saved as a runnable Turbine strategy, so any band configuration in the grid can be pulled up and inspected directly.

One note on the sweep size: because the grid varies two parameters across ten values each, there are 100 cells total, and this is the trial count the robustness checks assume. That matters for how much weight to put on any single winner.

Top results

The best cell was the widest band on the board: floor 0.01 / ceiling 0.99, producing $1,393.68 total PnL across 866 trades, a 65.9% win rate, ROI 2,322.81%, Sharpe 0.96, and max drawdown -78.36.

The runner-up, floor 0.36 / ceiling 0.99, landed at $1,351.02 PnL over 837 trades, 66.4% win rate, Sharpe 0.94, with the same -78.36 drawdown. Third place, floor 0.37 / ceiling 0.99, came in at $1,305.63 over 832 trades, Sharpe 0.97.

A few things stand out. The top of the leaderboard is dominated by a 0.99 ceiling — every top-8 variant uses it. Floor values drift upward across the ranks (0.01, 0.36, 0.37, 0.38, 0.39, 0.40, 0.42, 0.41), but the ceiling barely moves. Trade counts also fall monotonically as the floor rises, from 866 down to 810, which is intuitive: a higher floor excludes cheaper contracts, so fewer entries clear the band. Win rates stay remarkably tight, all between 65.5% and 66.4%.

The marginals tell a consistent story on the floor axis: mean PnL declines steadily from $1,012.93 at floor 0.01 to $675.12 at floor 0.44. The ceiling axis is less tidy. Mean PnL at ceiling 0.56 is $626.77, dips at 0.58 to $580.24, then rises steadily to $1,250.71 at ceiling 0.99. The best cell on the floor axis and the best cell on the ceiling axis are not independent — floor 0.01 / ceiling 0.99 is the winner on both marginals simultaneously, which is the main reason it tops the grid.

The deflated Sharpe is 0.9995, and the expected max Sharpe under selection noise is 0.2575. These are not directly comparable — one is a probability, the other a Sharpe-scale benchmark. The deflated Sharpe is favorable for that specific check alone. It does not establish profitability, and it does not substitute for out-of-sample evidence. The neighborhood degradation figure is 0.1178.

Bottom results

The weakest cells cluster at the tight end of the band. The absolute worst, floor 0.44 / ceiling 0.58, produced $424.59 PnL over 356 trades, a 56.6% win rate, Sharpe 0.59, and -71.65 max drawdown. Just above it: floor 0.44 / ceiling 0.56 at $429.02 (322 trades, 59.8% win rate, Sharpe 0.62), and floor 0.44 / ceiling 0.57 at $451.42 (333 trades, 58.8% win rate, Sharpe 0.65).

Three patterns are visible here. First, the tight bands trade far less: 322–386 trades versus 810–866 at the top. Second, win rates drop into the 56–60% range, roughly 6 points below the leaders. Third, the tight-band drawdowns are actually shallower (-55.51 to -71.65) than the leaders' -78.36 to -103.83. The tight band loses less on each trade and makes less in aggregate — a lower-activity, lower-amplitude profile.

The marginals confirm the ceiling story. Mean PnL at ceiling 0.56 is $626.77 and at 0.58 is $580.24; the maximum at ceiling 0.56 is $1,114.74, but the maximum within the tight-band cluster is only around $451. The floor marginal shows the same gradient: floor 0.44 averages $675.12 across all ceilings versus $1,012.93 at floor 0.01. Together, the two marginal curves point in the same direction — wider is better on this data.

One inconsistency worth flagging honestly: the floor axis marginal means and maxima decline smoothly as the floor rises, but the ceiling axis is non-monotonic at the tight end (0.56 > 0.57 > 0.58, then rising again), and the cell-level values in that region do not line up cleanly with the marginal averages. Some of this is likely boundary interaction at the tight band — contracts priced near the edges get cut off — but the exact cause of the non-monotonicity in the ceiling marginal is not resolved by this dataset. It could be a sample-size artifact, an edge effect, or something in how the band interacts with the entry filter, and this report does not claim to know which.

Conclusion

The sweep gives a clear directional answer to the thesis question: yes, the tradable band materializes in the results. Wide bands ($1,305–$1,394 PnL, 65–66% win rate) substantially outperform tight bands ($425–$451 PnL, 57–60% win rate) across all 100 cells. The floor axis is a clean monotonic gradient; the ceiling axis is directionally consistent at the extremes but noisier in the middle.

That said, three caveats deserve equal billing.

First, the performance gap is partly a trade-count effect, not purely an edge effect. Wide bands trade 2.3–2.7× more often. Per-trade quality differs — win rate is ~6 points higher in wide bands — but the bulk of the aggregate PnL difference is volume.

Second, the drawdown profile is worse at the top. The best-PnL variants carry -78 to -104 max drawdowns, while the worst-PnL variants carry -55 to -72. If drawdown tolerance is the binding constraint, the ranking inverts.

Third, and most importantly, the robustness checks do not clear the bar for calling this validated. The permutation test result was not available to this report, and without it the