KXETH15M Entry-Family Isolation Study

Short disclaimer: This is historical simulation research on past market data. It is not investment advice, not a recommendation, and tells you nothing certain about future results. Options and prediction markets can lose money quickly.

Intro / thesis

The saved ETH 15-minute stack on Kalshi (KXETH15M) contains four distinct entry families that normally run together. When they run together, you can't tell which one is actually doing the work. This batch takes each family apart, runs it alone, and holds everything else constant so the comparison is clean.

The four families are:

• early_entry — ETH-only. Price between 0.40 and 0.60, ETH change_5m and velocity_1m both above zero, ETH price above vwap_1h, and ema_12_1m above sma_20_1m. A mirror rule fires the NO side on the full negative set.
• triple_confirm — ETH, BTC, and SOL all agree on change_5m above 0.0010 (or all below -0.0010), with ETH price above both vwap_1h and ema_12_1m (or below both). Price band 0.35–0.65.
• accel_momentum — ETH change_5m beyond ±0.0015 plus velocity_1m agreeing in sign. No cross-asset confirmation. Price band 0.20–0.80.
• triple_accel — ETH, BTC, and SOL all agree on change_5m beyond ±0.0015 and on change_15m sign. Price band 0.20–0.80.

The fifth candidate is the full stacked baseline: all four gates live at once. Everything — exit logic, risk caps, loop cadence, entry size, spread gate — is shared across all five, so PnL, trade count, win rate, and drawdown line up directly.

The core question: does the stacked entry earn its complexity, or does one family carry the whole thing?

Variant and strategy explanation

Every candidate shares the identical shell:

• Venue and market: Kalshi, series_ticker: KXETH15M, crypto market.
• Loop: every 10 seconds.
• Risk: max_position 60, price_floor 0.01, price_ceiling 0.99. Other runtime checks may still block orders.
• Entry size: 50 contracts on every fill.
• Spread gate: spread <= 0.015 on every entry rule.
• Entry window: time_to_expiry <= 10m on every entry rule.

The exit stack is four rules, evaluated top to bottom, first match wins and ends the tick:

1. position_size <= 10 AND unrealized_pnl > 3 AND price between 0.10 and 0.90 → sell all.
2. position_size >= 11 AND position_size <= 20 AND unrealized_pnl > 5 AND price between 0.10 and 0.90 → sell all.
3. position_size >= 21 AND unrealized_pnl > 8 → sell all.
4. unrealized_pnl < -5 AND price between 0.15 and 0.85 → sell all.

Each candidate then adds only its own entry rules on top of that shell. Candidates 2–5 each carry one family's YES and NO entry rules. Candidate 1 carries all four families' entry rules stacked.

One caveat you should read before the numbers. Entries are size 50 in every family. That means on a fill, position_size is already 50, so exit tier 3 (position_size >= 21) is the only take-profit tier that can fire on a live position. Tiers 1 and 2 require position_size <= 10 and 11–20 respectively, which a 50-contract entry never produces. All four tiers are kept verbatim across every candidate for fidelity to the saved stack, but in practice tier 3 and the stop are the working exits. This is a property of the shared exit stack, not a difference between candidates, so it doesn't break comparability — but it does mean the take-profit side is coarser than the tier names suggest.

Every successful variant in this batch is saved as a runnable Turbine strategy, so you can load, inspect, and re-run any of them directly.

Top results

Note on the ordering: these five candidates finish within a fairly tight band on ROI, so "top" here is really about which one you'd want to run given how much drawdown and how much trade flow you're willing to accept. Rank 1 is the highest total PnL, but it is not the best candidate on every column.

The headline number: the full stacked baseline tops PnL at 881.93 over 30 days on 1118 trades, with a 49.2% win rate and a -277.28 max drawdown. That is the largest drawdown in the batch and the weakest Sharpe (0.33) among the four highest-PnL candidates.

The interesting result is triple_confirm only. It produces 835.78 PnL — within about 5% of the baseline — on 254 trades, roughly one quarter the activity. Its win rate is 0.598 and its max drawdown is -105.62, the smallest of the five by a wide margin. Its Sharpe of 0.67 is the best in the batch. If you care about drawdown-adjusted behavior rather than raw PnL, this is the number that stands out.

early_entry only lands at 852.66 PnL on 676 trades with a 0.512 win rate, a -260.11 drawdown, and a Sharpe of 0.41. It carries a large share of the baseline's trade flow and a large share of its drawdown.

accel_momentum only is close on PnL at 832.47, on 574 trades, win rate 0.523, drawdown -215.51, Sharpe 0.44.

Bottom results

The bottom of the table is triple_accel only, at 547.94 PnL / 913.23% ROI on 248 trades, win rate 0.597, drawdown -190.80, Sharpe 0.34. It is the lowest PnL in the batch by a clear margin — roughly 38% below the baseline — despite having a win rate comparable to triple_confirm and a trade count comparable to triple_confirm. Same number of trades as triple_confirm, similar win rate, much lower PnL.

That gap is worth flagging directly. triple_accel and triple_confirm have nearly identical trade counts (248 vs 254) and nearly identical win rates (0.597 vs 0.598), yet triple_confirm produces about 288 more PnL. With the exit stack fixed and the entry size fixed, that difference is not explained by the mechanics documented here. It is an unresolved discrepancy in the results. It could come from differences in the price levels at which each family enters, from differences in the positions' paths to the exit triggers, or from something else — but nothing in the shared ruleset accounts for it, and I am not going to guess. Treat it as a real observation that needs a follow-up run to explain, not as a settled property of the two signals.

Second from the bottom on PnL is accel_momentum at 832.47, though on the drawdown column it sits above early_entry and the baseline. Third from the bottom by PnL is triple_confirm — which, again, is the best candidate in the batch on drawdown and Sharpe. Ranking by PnL and ranking by drawdown-adjusted return put different candidates on top, and that is the useful thing to take away from this table.

Conclusion

The stacked baseline wins on raw PnL, but it does so with the largest drawdown and the weakest Sharpe of the top four. The ROI spread across the top four candidates is narrow — 1469.88% down to 1387.45% — while the trade counts vary by more than 4x and the drawdowns vary by roughly 2.6x. Choosing between them is less about which one made the most in this 30-day window and more about what trade frequency and drawdown profile you're willing to carry.

If you're optimizing for drawdown-adjusted return on this window, triple_confirm only is the standout: best Sharpe, smallest drawdown, and PnL within about 5% of the baseline on a quarter of the trades. That is a materially different operating profile than the stacked version.

If you're optimizing for total activity and don't mind the deeper