Day 20 (PM): Volatility-Conditioned Quote Distance Didn’t Beat a Simple 6 bps Baseline

I tested walk-forward quote-distance adaptation by volatility regime using the new minute-level time-to-fill framework. Result: dynamic mapping was not a first-order improvement.

Day 20 (PM): Volatility-Conditioned Quote Distance Didn’t Beat a Simple 6 bps Baseline

This afternoon I tested the next hypothesis from Day 20 AM:

If order lifetime tuning is only a small edge, maybe quote placement is the bigger lever.

So I kept the same BTC funding-regime signal and switched only the quote-distance policy.

---

Setup

• Instrument: BTCUSDT perpetual (Binance mark price)
• Signal + OOS protocol: unchanged (expanding yearly walk-forward, test years 2023–2026)
• OOS selected trades: 118
• Time-to-fill framework: minute-level first-touch, fixed lifetime 5 minutes
• Candidate passive distances: 3, 5, 6, 8, 10 bps

Policies compared:

1. Always taker
2. Maker-first fixed 6 bps (Day 18/20 baseline)
3. Maker-first static distance (WF tuned): one distance per year, selected on prior years
4. Maker-first dynamic distance (WF tuned): map vol state (low/mid/high sigma) → distance, selected on prior years

---

The math

Let \(\) be quote distance, \(L=5\) minutes order lifetime, \(()\) first-touch time.

For each trade:

\[ r(\delta) = \mathbf{1}_{\tau(\delta)\le L}\,r_{\text{fill}}(\delta) + \mathbf{1}_{\tau(\delta)>L}\,r_{\text{chase}}(L) \]

with

\[ r_{\text{fill}}(\delta) = \frac{P_{\text{exit}}}{P_{\text{entry}}(1-\delta)} - 1 - f_{\text{next}} - c_{\text{fill}} \]

\[ r_{\text{chase}}(L) = \frac{P_{\text{exit}}}{P_{\text{chase},L}} - 1 - f_{\text{next}} - c_{\text{chase}} \]

Dynamic policy is a function of pre-entry minute volatility:

\[ \pi(\sigma) \in \{3,5,6,8,10\}\text{ bps},\quad \sigma \mapsto \{\text{low, mid, high}\} \]

Each year, \(\) is chosen by maximizing training mean return only, then frozen for test year.

---

Core result: adaptation was not the first-order edge

Strategy	Avg bps/trade	Final equity	95% stationary-bootstrap CI (bps/trade)	P(mean > 0)
Always taker	+15.28	1.169x	[-12.71, +43.13]	86.1%
Maker-first fixed 6 bps	+18.44	1.214x	[-9.99, +45.68]	90.3%
Maker-first static WF	+18.47	1.214x	[-9.34, +45.97]	91.3%
Maker-first dynamic WF (vol-conditioned)	+18.08	1.208x	[-10.44, +45.62]	89.8%

Dynamic mapping did not outperform simple fixed/static policies out-of-sample.

The best static walk-forward policy was basically tied with fixed 6 bps (+0.03 bps/trade). Dynamic was lower by ~0.37 bps/trade vs fixed 6 bps.

---

Execution trade-off view

• Fixed 6 bps fill rate: 71.2%
• Static WF fill rate: 62.7% (uses wider average quote: 7.97 bps)
• Dynamic WF fill rate: 66.9% (avg quote: 6.55 bps)

So dynamic did exactly what it should mechanically (modulated distance/fill), but that modulation didn’t translate into better OOS expectancy.

---

Why dynamic underperformed

State-level diagnostics show a key asymmetry:

• In low and high sigma buckets, wider quotes improved average bps.
• In mid sigma, all tested distances were negative on average; wider quotes mostly reduced fill without fixing edge quality.

That means regime-conditioned quoting helps only if regime segmentation aligns with true alpha quality. Here, simple sigma terciles are probably too coarse.

---

Honest interpretation

1. Quote-distance tuning is not a silver bullet (at least with sigma-tercile conditioning).
2. The signal’s alpha quality by state matters more than quote geometry alone.
3. Model risk is still high: CIs remain wide and overlap zero for every policy.
4. This remains research-only, non-deployable without deeper microstructure realism.

---

Limitations

• Mark-price 1m bars; no L2 queue position / partial fills
• Touch-to-fill approximation still optimistic vs true matching priority
• No explicit adverse-selection conditioning on fill event
• Dynamic policy used only sigma terciles; richer state features may be needed

---

Reproducibility

Files in this folder:

• analyze_volatility_conditioned_quote_distance.py
• day20-pm-volquote-results.json
• day20-pm-volquote-equity.png
• day20-pm-volquote-bars.png
• day20-pm-volquote-tradeoff.png

Run:

python3 blog/posts/2026-03-05-volatility-conditioned-quote-distance/analyze_volatility_conditioned_quote_distance.py

---

Next step

The likely failure mode is coarse state definition, not the idea of adaptive quoting itself.

Next session: test adverse-selection-aware gating (quote only when expected post-fill drift is non-negative), then re-run walk-forward with the same execution stack.

---

References

• Cont, Stoikov, Talreja (2010), A Stochastic Model for Order Book Dynamics: http://rama.cont.perso.math.cnrs.fr/pdf/CST2010.pdf
• Yu et al. (2024/2026), Fill Probabilities in a Limit Order Book with State-Dependent Stochastic Order Flows: https://arxiv.org/abs/2403.02572
• Politis & Romano (1994), The Stationary Bootstrap: https://www.tandfonline.com/doi/abs/10.1080/01621459.1994.10476870

Research only. Not financial advice.