Day 26: Lifetime Tuning by Regime
Summary
Result: NOT DEPLOYABLE — Marginal improvement, adaptive doesn’t help.
| Strategy | Expected PnL | Improvement |
|---|---|---|
| Fixed 5min | 6.39 bps | baseline |
| Fixed 10min | 7.91 bps | +1.52 bps |
| Fixed 15min | 8.37 bps | +1.98 bps |
| Vol-adaptive | 7.57 bps | +1.18 bps |
Verdict: Longer lifetime helps marginally (+0.46 bps from 10min→15min), but adaptive regime-switching doesn’t add value.
The Question
We’ve optimized quote distance (Day 25). Now: does order lifetime need to adapt to market conditions?
- High volatility: market moves fast, shorter lifetime reduces adverse selection exposure
- Low volatility: stable market, longer lifetime improves fill rate
- Volume: higher volume = faster fills
Methodology
Model
Fill probability increases with: - Longer lifetime - Lower volatility - Higher volume ratio
PnL calculation includes real post-fill drift (adverse selection cost): - If we get filled at time t, measure actual price movement in next N minutes - Longer lifetime = more time for price to move against us
Walk-Forward
- Data: 5-minute candles, Jan 2022 – Dec 2024
- Train: 2022-2023
- Test: 2024-2025 (157,895 samples)
- Quote distance: 10 bps (from Day 25 optimization)
Strategies Tested
- Fixed 5min: baseline
- Fixed 10min: longer lifetime
- Fixed 15min: longest tested
- Vol-adaptive: 5min (high vol) → 10min (mid) → 15min (low vol)
Results
| Strategy | Expected PnL | vs 5min | vs 10min |
|---|---|---|---|
| Fixed 5min | 6.39 bps | — | - |
| Fixed 10min | 7.91 bps | +1.52 bps | — |
| Fixed 15min | 8.37 bps | +1.98 bps | +0.46 bps |
| Vol-adaptive | 7.57 bps | +1.18 bps | -0.34 bps |
Key Findings
- Longer lifetime wins: 15min > 10min > 5min
- More time = higher fill rate = more expected value
- Adverse selection cost doesn’t fully offset fill gains
- Adaptive doesn’t help: vol-adaptive actually underperforms fixed 10min
- Regime switching adds noise, not signal
- Volatility prediction at 5-15min horizon isn’t reliable enough
- Marginal improvement: +0.46 bps (10min → 15min) is within noise
- Not statistically significant
- Not worth the complexity
Conclusion
Not deployable. The improvement from 10min → 15min lifetime is marginal (+0.46 bps) and within noise. Adaptive strategies actively hurt performance.
Practical recommendation: Stick with 10-minute lifetime (or whatever current baseline is). The complexity of regime-adaptive lifetime doesn’t pay off.
Files
- research_full.py — Full backtest with real drift
- backtest_results.csv — Results data
- verdict.json — Decision record