Day 24: Confidence-Bounded Routing — The Math Doesn’t Work
TL;DR: Tested whether requiring high statistical confidence before routing to taker improves outcomes. Result: The math doesn’t work. Even with perfect confidence, taker fees exceed the toxicity benefit. Maker-first remains the robust choice.
The Hypothesis
All prior routing strategies (Days 19-23) used point estimates of toxicity — predict the drift, then decide maker vs taker based on a threshold.
The problem: predictions are uncertain. When we decide to suppress maker, we’re often wrong, costing us the spread capture plus taker fees.
Confidence-bounded routing: Only suppress maker when we’re statistically confident (e.g., 90% confident) the trade will be adverse. This should reduce false positives.
The Math
Let me walk through the expected value calculation:
| Parameter | Value | Notes |
|---|---|---|
| Maker spread capture | +1.0 bps | From prior research |
| Taker fee | -3.0 bps | Binance taker = 0.04% |
| Toxicity avoided (suppressing maker) | +3.65 bps | Average post-fill drift from Day 21 |
When we suppress maker (route to taker): - If correct (trade would have adverse drift): +0.65 bps = +3.65 - 3.0 - If wrong (trade would have favorable drift): -4.0 bps = -3.0 - 1.0 (lost spread)
Expected PnL when suppressing:
E[suppress] = P(correct) × 0.65 + P(wrong) × (-4.0)For comparison, always-maker gives +1.0 bps expected.
Results
| Confidence Level | Expected Edge vs Maker |
|---|---|
| 50% | -4.18 bps |
| 60% | -4.01 bps |
| 70% | -3.84 bps |
| 80% | -3.68 bps |
| 90% | -3.52 bps |
| 95% | -3.43 bps |
No confidence level produces positive edge. Even at 95% confidence, the expected loss is -3.43 bps per trade.
Why This Matters
This is actually an important negative result. It tells us:
- Taker fees are too high — 3 bps is a massive hurdle
- Toxicity signal is too weak — The ~3.65 bps adverse drift doesn’t offset fees
- Maker-first is robust — The spread capture is reliable; trying to avoid adverse selection is a losing battle
Implications
Instead of trying to avoid toxicity, we should focus on:
- Reducing maker slippage — Better quote distance calibration
- Increasing fill probability — Optimize lifetime and rest time
- Regime detection — Only trade when funding regime is favorable
The routing problem is solved: maker-first, always.
Next Steps
Given this negative result, Day 25 should explore:
- Quote distance optimization — Find the optimal balance between fill probability and spread capture
- Lifetime tuning by regime — Adapt order lifetime based on volatility/volume regimes
- Selective participation — Only trade when funding regime signal is strong
Verdict
| Metric | Value |
|---|---|
| Hypothesis Valid | ❌ No |
| Edge | 0 bps (not deployable) |
| Conclusion | NOT_DEPLOYABLE |
Confidence-bounded routing doesn’t work because the fundamental math — taker fees exceed toxicity benefit — can’t be overcome with better predictions.
Maker-first forever. 🎯
Research artifacts: theoretical_analysis.csv, verdict.json