Kalshi variant 005 led the family with 2.90% ROI. The weakest completed variant

Short disclaimer This is a simulation study of historical data only. It does not forecast future returns. All results are backward-looking and may not persist.

Intro / thesis In Kalshi’s KXHIGHNY daily high-temperature markets, the headline NWS forecast is widely watched. But the forecast itself is a single number issued early in the day. What changes intraday is the observed temperature. The core idea tested here is that you improve your edge by trading the gap between current KLGA temperature and that static forecast high, not by reacting to the forecast alone. Buy YES when the station is already within a couple of degrees of the forecast high and the next three hours are not cooling; buy NO when the station is still well below forecast, the near-term hourly forecast shows no recovery, and the price is elevated. The research also asked: across risk bounds, loop cadence, position sizing, and price filters, does this forecast-vs-observation framework outperform a forecast-only approach?

Variant and strategy explanation The study starts from a baseline DSL and varies it systematically. The baseline monitors KLGA (LaGuardia) via NWS, refreshing every 60 seconds on a 60-second loop. It computes temp_delta_from_forecast_high_f, the difference between current temperature and the NWS forecast high. It also pulls temp_hour_3, a three-hour-ahead temperature estimate.

• Buy YES rule: Trigger when temp_delta_from_forecast_high_f ≥ −2.5°F (observation already close to forecast high), temp_hour_3 ≥ current temperature (no imminent cooling), wind ≤ 20 mph, price ≤ 0.70, and at least 2 hours until expiry. Position size: 3 contracts.
• Buy NO rule: Trigger when temp_delta_from_forecast_high_f ≤ −7.0°F (observation well below forecast), temp_hour_3 ≤ current temperature (no recovery), price ≥ 0.30, and at least 2 hours until expiry. Position size: 3 contracts.
• Exits: Sell all if unrealized PnL ≥ +4.0, or ≤ −5.0, or if expiry is 30 minutes away.
• Risk: Max position 10, price floor 0.10, price ceiling 0.90, max loss 6.0.

The 100 variants perturb parameters around this base: delta thresholds, time-to-expiry gates, price filters, wind thresholds, position sizing, exit PnL levels and loop cadence. Each variant is a runnable Turbine strategy, saved and identified by slug.

Top results The best-performing variants were extremely consistent in structure, clustering around tiny positive edges.

• Rank 1–4: Identical realized performance. Max drawdown −4.12, total PnL +0.29, ROI +2.9%, win rate 16.7% on exactly 30 trades. Sharpe ratio effectively zero. These variants (005, 006, 007, 008) slightly tweak parameters from the base without materially changing outcomes — the equity curve is flat-to-slightly-positive with shallow, short-lived drawdowns.
• Rank 5–8: A second cluster with max drawdown −2.07, total PnL +0.12, ROI +2.4%, 18 trades, same 16.7% win rate. These variants (001–004) took fewer trades but produced a slightly gentler drawdown profile.

In all top variants, the strategy never loses heavily and never wins large. It scrapes a tiny positive PnL by avoiding catastrophic “forecast-chasing” entries and using contrarian entries only when the observation-vs-forecast gap is materially supportive. The theme: filtering for strong alignment between current observation, the three-hour temperature trend, and price keeps the strategy from bleeding, even though it rarely captures a large move.

Each successful variant has been saved as a runnable Turbine strategy for possible further inspection.

Bottom results The worst variants fall into two failure modes.

• Severe over-trading with relaxed edge filters: Variants 068, 073, 085, 086, 088. Max drawdowns range from −10.43 to −18.01, total PnL from −10.00 to −17.16, zero win rate on up to 50 trades. These variants lowered or removed near-term temperature trend requirements, accepted windier conditions, or bought YES at a wider delta — allowing the strategy to enter on marginal setups. The result was a rapid series of small losses that compounded into large drawdowns.
• Overly tight exit and sizing that erodes edge: Variants 081, 082, 083 posted max drawdown −5.97, total PnL −5.75, ROI −115% on 18 trades with zero wins. They combined restrictive stop-loss levels or price ceilings that forced exits before the underlying gap could revert.

The common thread among bottom variants is degrading the “observation-vs-forecast” signal purity. Weakening the delta threshold, ignoring the hourly temperature direction, or letting the strategy trade at poor prices universally harmed performance.

Conclusion Within this historical simulation, a forecast-vs-observation gap strategy for KXHIGHNY yields slightly positive median outcomes only when it adheres tightly to signal discipline. The top tier produced a thin positive edge with shallow drawdowns, while any relaxation of the delta and trend filters led to consistent, accelerating losses. The forecast-headline only approach was implicitly tested as a weaker baseline in many bottom variants, and it performed materially worse. The gap signal is fragile — it helps avoid bad trades rather than find great ones. This research makes no claim about future viability; it simply observes that, in the tested historical window, a disciplined observation-vs-forecast gap filter was the necessary condition for non-negative outcomes.

Long disclaimer This report presents results from a historical simulation conducted for research purposes. No actual trading was performed, and past simulated performance does not guarantee future results. The analysis uses market data and weather data available at the time of the study; data quality, latency, and execution assumptions differ from live conditions. Turbine strategies are mechanical DSL rules and do not adapt to regime changes. Any strategy, including those saved from this study, may produce losses in live markets. This report is not investment advice, a trade recommendation, or an offer to buy or sell any financial instrument. Use the information at your own risk, and consult a qualified professional before engaging in prediction-market trading.