SPXW 0DTE / Selection Bias / Full-Population Validation
Donchian Breakout on SPX 0DTE — A Negative-Result Case Study
Research completed / statistically significant loss
A four-year out-of-sample evaluation of a Donchian 5-minute breakout on SPX 0DTE options. Full-population Databento tick validation across three exit policies gives a real-executable PnL of roughly −$107K to −$117K per year, with bootstrap 95% confidence intervals entirely below zero.
Full materials
Research Question
Does a one-trade-per-day Donchian 5-minute breakout signal on SPX produce statistically distinguishable positive expectancy when expressed as long 0DTE option positions, after realistic execution friction measured directly from tick data?
The strategy is deliberately minimal: a 30-bar 5-minute Donchian channel, an at-the-money 0DTE strike with 5-point granularity, at most one entry per day, and three candidate exit policies (always_in, reversal_off, reversal_close_only).
Methodology
- Evaluate over a four-year out-of-sample window (2022-05 to 2026-05).
- Download Databento OPRA tick data for every trade in all three exit-policy ledgers within the coverage window — no stratified sample, no per-tier ratio, no extrapolation.
- Compute real-executable PnL directly, per trade, per leg.
- Bootstrap-resample to put confidence intervals around the annualized PnL.
- Cross-check the underlying SPX delta-1 PnL to separate signal edge from option-layer friction.
Result
The result is unambiguously negative for all three policies. Full-population direct measurement on 3,146 strictly-validatable trades gives roughly −$107K/year (reversal_off, the least bad) to −$117K/year (always_in), with bootstrap 95% confidence intervals entirely below zero. The underlying SPX-level signal is also not significantly positive, so switching to a linear vehicle (ES futures, SPY ETF) would not rescue it.
Why It Matters
Earlier drafts (v1–v4) reported successively smaller positive headlines (+$80K → +$50K → +$32K) that did not survive full-population measurement. Two compounding biases produced them: Black-Scholes priced with VIX1D under-prices ATM 0DTE entry premium by roughly 85%, and a strict-filter stratified sample silently excluded expired-worthless trades (27% of the population, all losers). The lasting value of this work is the methodological protocol — full-population direct validation — and the documented selection-bias mechanism, not a deployable strategy.