Conformal Tail-Coverage Slippage Playbook

Turning Slippage Uncertainty from “Model Confidence” into a Controlled Risk Budget

Why this note: Many execution stacks optimize expected slippage but under-govern uncertainty drift. When regime shifts hit (latency, liquidity, venue behavior), intervals become miscalibrated and tactic selection silently overcommits. Conformal calibration gives a practical way to keep uncertainty coverage honest in production.

1) Failure Mode in One Sentence

Your slippage model can keep good average error while losing tail coverage, causing over-aggressive routing exactly when uncertainty is highest.

2) Cost Objective (Add Miscalibration Penalty)

For child-order action (a), score with:

[ J(a|x) = \mathbb{E}[C_{IS}|x,a] + \lambda,\text{CVaR}_{q}(C_{IS}|x,a) + \eta,\text{MissRisk}(x,a) + \gamma,\text{CalibGap}(x) ]

Where:

• (\mathbb{E}[C_{IS}]): expected implementation shortfall
• (\text{CVaR}_{q}): tail slippage penalty (e.g., q=95%)
• (\text{MissRisk}): completion/deadline risk
• (\text{CalibGap}): empirical coverage error of prediction intervals

Most stacks include first 2–3 terms. The missing term is often calibration health.

3) Core Interval Construction (Deployable)

Start with any quantile model producing ([\hat q_{\alpha}(x), \hat q_{1-\alpha}(x)]). On a calibration slice, compute nonconformity:

[ r_i = \max{\hat q_{\alpha}(x_i)-y_i,; y_i-\hat q_{1-\alpha}(x_i)} ]

Let (\hat\tau) be the ((1-\varepsilon))-quantile of (r_i). Final interval:

[ \mathcal{I}(x)=\big[\hat q_{\alpha}(x)-\hat\tau,;\hat q_{1-\alpha}(x)+\hat\tau\big] ]

Interpretation: quantile model gives local shape; conformal correction restores finite-sample coverage discipline.

4) Telemetry That Matters in Live Routing

Track by symbol × venue × liquidity regime:

1. Empirical Coverage Gap (ECG) [ ECG = \hat p_{cover} - (1-\varepsilon) ]

2. Tail Undercoverage Rate (TUR)

   • Fraction where realized slippage exceeds upper bound.
   • Split by latency and spread regimes.

3. Interval Width Efficiency (IWE) [ IWE = \frac{\text{median interval width}}{\text{median realized |error|}} ] Too high = conservative/slow; too low + TUR↑ = dangerous.

4. Decision Boundary Breach Rate (DBBR)

   • Cases where chosen action assumed safe ex-ante but upper interval exceeded control threshold ex-post.

5. Coverage Drift Velocity (CDV)

   • Short-horizon slope of ECG/TUR (e.g., 30–60 min rolling).

If TUR↑ + DBBR↑ while mean error looks stable, your router is “confidently wrong.”

5) Data Contract (Point-in-Time Safe)

At decision time per child:

• decision_ts, send_ts, ack_ts, fill_ts
• action (passive/pegged/IOC/cross), venue, size, urgency_state
• pred_mean_cost, pred_q05, pred_q50, pred_q95
• conformal_tau, final_lo, final_hi
• realized_cost
• market_state (spread, imbalance, depth slope, volatility, book age)
• infra_state (queue delay, packet-loss proxy, clock quality)
• calibration_bucket_id (so coverage can be audited by bucket)

Critical: log both pre-conformal and post-conformal intervals; otherwise you can’t diagnose whether failure is model quality or calibration failure.

6) Regime-Aware Extensions

A) Weighted conformal under covariate shift

When live feature distribution differs from training, use weights (density-ratio style) so calibration focuses on current regime.

B) Sequential/time-series conformal refresh

Use rolling recalibration windows to reduce stale interval error in intraday regime flips.

C) Bucketed conformal

Maintain separate (\hat\tau) by liquidity/latency buckets (not one global scalar), with minimum sample guards.

7) Execution Controller (GREEN → SAFE)

• GREEN: Coverage in-band, widths efficient. Normal policy.
• AMBER_CALIB: ECG drifting, TUR rising mildly. Increase risk aversion, reduce passive confidence, tighten child-size caps.
• RED_CALIB: TUR/DBBR breach hard limits. Switch to completion-safe tactics, reduce venue experimentation, increase urgency penalties.
• SAFE_FALLBACK: Calibration unreliable (sample collapse or rapid shift). Use conservative deterministic policy until recalibration recovers.

Use hysteresis and minimum dwell time to avoid oscillation.

8) Practical Threshold Examples

• (|ECG| > 2.0%) for 3 windows → AMBER
• TUR (upper) > target + 1.5% and CDV positive → RED
• DBBR > 5% in high-urgency orders → immediate RED
• Calibration sample count below floor in active bucket → SAFE_FALLBACK

Tail health, not RMSE, should gate policy aggression.

9) Rollout Plan

1. Shadow (2 weeks): compute conformal intervals and TUR/DBBR only.
2. Paper controller: simulate GREEN/AMBER/RED transitions against historical streams.
3. Canary (low-notional symbols): enable calibration-aware action scoring.
4. Promotion gates: improve p95/p99 IS and completion reliability, with no hidden over-conservatism blowout.
5. Runbook drills: force synthetic covariate-shift scenarios and verify SAFE_FALLBACK behavior.

10) Fast Checklist

[ ] Log pre/post-conformal intervals for every child decision
[ ] Monitor ECG, TUR, IWE, DBBR, CDV by symbol×venue×regime
[ ] Add calibration penalty into action objective
[ ] Use rolling/bucketed recalibration with sample floors
[ ] Deploy GREEN/AMBER_CALIB/RED_CALIB/SAFE_FALLBACK states
[ ] Gate rollout on tail metrics + completion, not mean error

References

• Romano, Patterson, Candès (2019), Conformalized Quantile Regression (NeurIPS; arXiv:1905.03222).
• Tibshirani et al. (2019), Conformal Prediction Under Covariate Shift (weighted conformal framework).
• Xu & Xie (2021), Conformal Prediction Interval for Dynamic Time-Series (ICML, EnbPI).
• Almgren & Chriss (2000), Optimal Execution of Portfolio Transactions (execution-cost control baseline).

TL;DR

Slippage control fails when uncertainty intervals are uncalibrated, not only when point forecasts are wrong. Conformalized tail coverage gives a simple production mechanism to keep interval promises honest, throttle aggression during drift, and protect p95/p99 execution outcomes.