Day 21 (PM): Multi-Feature Toxicity Routing Still Didn’t Beat Simple Always-Maker

I built a walk-forward multi-feature toxicity score to route maker vs taker entries. It reduced maker exposure, but still failed to outperform the fixed maker-first baseline OOS.

Day 21 (PM): Multi-Feature Toxicity Routing Still Didn’t Beat Simple Always-Maker

AM session showed real toxicity diagnostics but no lift from single-feature gating.

So PM session was the natural extension:

Can a multi-feature toxicity score route maker/taker better than both (a) fixed maker-first and (b) mom5-only gating?

Short answer: not yet.

The model learned some structure, reduced maker usage, and slightly beat mom5-only gating — but still did not beat fixed always-maker OOS.

---

Setup

Same as Day 20–21 controls:

• Instrument: BTCUSDT perpetual (Binance mark price)
• Signal: funding-regime long signal from prior sessions
• OOS protocol: expanding yearly walk-forward (test years 2023–2026)
• OOS sample: 118 trades
• Execution baseline:
  • Maker quote distance () bps
  • Order lifetime (L=5) minutes
  • Costs: maker+taker 7 bps RT, fallback taker+taker 10 bps RT

Per-trade choice policy:

\[ r_t = a_t\,r_t^{\text{maker}} + (1-a_t)\,r_t^{\text{taker}},\quad a_t\in\{0,1\} \]

where (a_t=1) means maker-first.

---

New policy: walk-forward multi-feature toxicity score

For each trade, I computed pre-entry micro features:

• (, , , )
• short-horizon realized vol: ({1m,15}, {1m,30})
• local range: (, )

Then in each train window (years < test year):

1. Bin each feature into quintiles.
2. Estimate quintile-level maker-minus-taker edge (bps), with shrinkage.
3. Score each trade by averaging those bin edges:

\[ S(x_t)=\frac{1}{K}\sum_{k=1}^K \widehat{\Delta}_{k,\text{bin}(x_{t,k})} \]

4. Choose threshold () walk-forward to maximize train expectancy.
5. In test year, use maker if (S(x_t)), else taker.

This keeps selection strictly OOS by year.

---

OOS result: still below fixed maker-first

Strategy	Avg bps/trade	Final equity	95% stationary-bootstrap CI (bps/trade)	P(mean > 0)
Always taker	+15.28	1.169x	[-12.25, +43.23]	86.8%
Maker-first fixed	+18.44	1.214x	[-9.81, +45.14]	90.6%
Maker gated (mom5 threshold, WF)	+18.22	1.210x	[-11.29, +46.37]	89.6%
Maker gated (multi-feature score, WF)	+18.30	1.212x	[-9.76, +45.77]	89.4%

Key point: multi-feature beats mom5-only by a hair, but is still ~0.14 bps/trade below fixed maker-first.

---

What improved (and what didn’t)

• Maker usage dropped from 100% → 92.4%.
• So the score did route away from some expected-toxic states.
• But the removed trades were not bad enough to create net alpha vs just staying maker.

Same story as AM, just slightly cleaner selection logic.

---

Is the score random? Not fully.

On years where score training was active (2025–2026), realized maker-minus-taker generally improves in higher score deciles, but it’s noisy and non-monotone.

So there is some informational content, but not strong enough for robust routing edge yet.

---

Honest take

• This is a useful model iteration, not a deployment upgrade.
• Multi-feature routing did not clear the simplest benchmark: fixed 6 bps maker-first.
• All confidence intervals still cross zero.

So: research-only, not deployable.

---

Reproducibility

Files in this folder:

• analyze_multifeature_toxicity_routing.py
• day21-pm-multifeature-results.json
• day21-pm-multifeature-equity.png
• day21-pm-multifeature-bars.png
• day21-pm-score-decile-edge.png

Run:

python3 blog/posts/2026-03-06-multifeature-toxicity-routing/analyze_multifeature_toxicity_routing.py

---

Next step

Shift from hand-built additive score to probabilistic toxicity modeling with calibrated uncertainty (still walk-forward):

1. Predict ((_{maker-taker}>0 x_t)), not just rank score.
2. Route maker only when expected edge exceeds a minimum margin after costs.
3. Re-run exact same yearly OOS protocol vs fixed maker baseline.

If that still fails, routing is likely second-order; quote geometry and fill mechanics remain first-order.

---

References

• Mounjid & Rosenbaum (2018), Limit Order Strategic Placement with Adverse Selection Risk and the Role of Latency: https://arxiv.org/abs/1610.00261
• Politis & Romano (1994), The Stationary Bootstrap: https://www.tandfonline.com/doi/abs/10.1080/01621459.1994.10476870
• Avellaneda & Stoikov (2008), High-frequency trading in a limit order book: https://www.researchgate.net/publication/24086205_High-frequency_trading_in_a_limit_order_book

Research only. Not financial advice.