Hierarchical Bayesian Cross-Symbol Slippage Transfer Learning Playbook

Why this matters

In live execution, most symbols are data-poor while a few liquid names are data-rich. A single global slippage model underfits symbol-specific behavior, and one-model-per-symbol is too noisy for thin names.

A hierarchical Bayesian setup gives a practical middle path:

• Global structure for stability
• Symbol-level adaptation for realism
• Uncertainty-aware decisions for risk control

This is especially useful when launching new symbols, new venues, or new tactics with limited local history.

1) Problem setup

Target (per parent order or slice):

• y = realized slippage_bps relative to arrival/mid benchmark

Core predictors:

• Participation (qty / interval_volume)
• Spread (bps), depth, queue imbalance
• Volatility (short-horizon realized vol)
• Time bucket (open/mid/close/auction proximity)
• Urgency, tactic type, side, venue
• Regime flags (stress/news/VI/circuit/auction)

Challenge:

• Heavy tails, heteroskedasticity, and symbol regime drift
• Sparse observations for many symbols

2) Model architecture (practical)

Use a hierarchical location-scale model with robust likelihood.

[ \begin{aligned} y_i &\sim \text{StudentT}(\nu, \mu_i, \sigma_i) \ \mu_i &= X_i\beta_{g(i)} + Z_i\gamma_{s(i)} \ \gamma_s &\sim \mathcal N(0, \Sigma_\gamma) \end{aligned} ]

Where:

• g(i) = group (e.g., venue × tactic × cap bucket)
• s(i) = symbol
• β_group: partially pooled group coefficients
• γ_symbol: symbol random effects (intercept + selected slopes)
• Student-t handles outliers/tail events better than Gaussian

For heteroskedasticity:

[ \log \sigma_i = W_i\alpha_{g(i)} + u_{s(i)},\quad u_s \sim \mathcal N(0, \tau_u^2) ]

This gives both expected slippage and uncertainty conditioned on context.

3) Transfer-learning logic

3.1 Pooling strategy

Build a hierarchy like:

1. Global
2. Market bucket (KOSPI/KOSDAQ, venue)
3. Liquidity tier
4. Symbol

A new or sparse symbol inherits higher-level priors. As data arrives, posterior naturally shifts toward symbol-specific behavior.

3.2 Feature sharing

Share nonlinear transforms globally:

• sqrt(participation) (impact-like)
• spread * participation
• vol * participation
• queue_imbalance * side

Let symbol random effects adjust sensitivity, not redefine the full model.

3.3 Cold-start policy

For symbols with < N_min observations:

• Use posterior predictive from upper levels
• Apply tighter participation caps
• Increase uncertainty multiplier in scheduler

4) Data spec and labeling

Minimum grain:

• Parent order id, slice id, timestamps
• Intended vs executed qty/price
• Market snapshots at decision + fill times
• Venue/tactic metadata
• Reject/cancel/partial-fill path

Label hygiene:

• Keep censored outcomes (unfilled remainder) as explicit states
• Separate benchmark definitions: arrival, decision-mid, interval-VWAP
• Attach event flags (news, auction, VI, halts)

Recommended splits:

• Rolling time split (no leakage)
• Stress-window holdout for tail validation
• New-symbol holdout for transfer-learning check

5) Fitting and online update

Offline (daily/weekly)

• Fit full hierarchical model (Stan/PyMC/TFP)
• Save posterior summaries + calibration diagnostics
• Export compact runtime artifacts (means, covariance, quantile maps)

Online (intra-day)

• Bayesian updating for intercept/risk scale via lightweight filters
• CUSUM/change-point monitor triggers prior inflation on regime breaks
• Keep hard guardrails independent of model confidence

Pseudo-flow:

1. Score context → posterior predictive (mean, p90, p99)
2. Compute expected cost + risk penalty
3. Choose tactic/participation under risk budget
4. Observe realized outcome
5. Update lightweight state + drift monitors

6) Decision policy integration

Use uncertainty-aware control, not point estimate only.

Example objective:

[ \text{score} = \mathbb E[y] + \lambda \cdot \text{TailRisk}_{q} + \eta \cdot \text{MissRisk} ]

• TailRisk_q: predictive quantile (e.g., p95/p99)
• MissRisk: probability of not finishing within horizon

Policy ladder:

• Green: low mean + low tail → normal participation
• Yellow: moderate tail → reduce clip, increase patience
• Red: high tail or high epistemic uncertainty → defensive mode (smaller child orders, wider randomization, optional pause)

7) Evaluation metrics (must-have)

Accuracy:

• MAE/RMSE for conditional mean
• Pinball loss (p50/p90/p95)

Calibration:

• Coverage of predictive intervals (e.g., 90% target)
• PIT histogram / reliability curve

Economic impact:

• Realized IS reduction vs baseline
• Tail-loss reduction (p95/p99 slippage)
• Completion-rate stability under stress

Transfer quality:

• New-symbol performance after k trades
• Regret vs per-symbol-only and global-only baselines

8) Failure modes and mitigations

1. Over-pooling (ignoring symbol uniqueness)

   • Add random slopes for critical features
   • Relax shrinkage priors for high-liquidity outliers

2. Under-pooling (too noisy)

   • Increase prior strength for sparse symbols
   • Collapse unstable subgroup hierarchy

3. Regime discontinuity

   • Change-point detector + fallback conservative policy
   • Time-decayed likelihood weighting

4. Label contamination

   • Strict benchmark versioning
   • Fill/cancel state machine audit

9) Implementation blueprint (Vellab-friendly)

Phase 1 (1 week)

• Define canonical slippage dataset contract
• Build global + symbol random-intercept model
• Produce posterior predictive API: mean/p90/p99

Phase 2 (1–2 weeks)

• Add random slopes (participation, spread, vol)
• Add heteroskedastic head (log sigma)
• Integrate uncertainty-aware scheduler controls

Phase 3 (ongoing)

• Drift/change-point auto-monitoring
• Champion–challenger with regret budget
• Online recalibration and rollback rules

10) Practical defaults

• Likelihood: Student-t (ν learnable, lower-bounded)
• Priors: weakly informative, scale-normalized features
• Refit cadence: daily + emergency refit on structural breaks
• Runtime: quantile-oriented outputs first, not just mean
• Safety: hard participation/price-band caps outside model

Bottom line

Hierarchical Bayesian slippage modeling is a strong production pattern when you need both:

1. Cross-symbol transfer (to avoid cold-start blindness), and
2. Symbol-level realism (to avoid one-size-fits-none errors).

Treat it as a decision system (prediction + uncertainty + guardrails), not a pure forecasting exercise.