adverse-selection-classifier

name	adverse-selection-classifier
description	Predict informed vs noise trades for dynamic spread/kappa adjustment. Use when building AS prediction, debugging fill losses, adding liquidation cascade detection, or tuning toxic flow response. Covers trade classification, 26-feature engineering, MLP classifier, and real-time integration.
requires	["measurement-infrastructure"]
user-invocable	false

Adverse Selection Classifier Skill

Purpose

Predict which incoming trades are informed before they hit you, enabling dynamic spread/kappa adjustment to protect against toxic flow. This is often the highest-value model for PnL improvement because adverse selection is typically the largest drag on market maker profits.

When to Use

• Building adverse selection prediction from scratch
• Debugging why you're losing money on fills
• Adding liquidation cascade detection
• Tuning when to widen spreads vs stay tight
• Any time adverse selection is your biggest PnL drag

Prerequisites

• measurement-infrastructure with post-fill price tracking
• signal-audit to identify predictive features
• Historical trades with ex-post price evolution

---

Core Concept: Trade Classification

Every trade comes from a latent source:

1. Noise traders: No information, mean-reverting flow
2. Informed traders: Have directional information, toxic to market makers
3. Liquidations: Forced selling/buying, extremely toxic but predictable
4. Arbitrageurs: Cross-exchange spread closure, predictable timing

You can't observe the true type, but you can:

• Infer it from ex-post price movement (for training)
• Predict it from observable features (for real-time use)

---

Labeling Strategy

Since true intent is unobservable, label based on ex-post outcomes:

Labels: Informed, Noise, Liquidation, Arbitrage. Priority: check liquidation context first (OI drop >2% + extreme funding), then arbitrage (cross-exchange spread >5bps closed within 500ms), then informed vs noise based on ex-post 10s price move (>5bps in trade direction = informed).

See implementation.md for full labeling code.

Label Distribution (Typical)

Label	Frequency	Toxicity
Noise	60-70%	Low (slightly positive expectation)
Informed	20-30%	Medium (negative expectation)
Arbitrage	5-10%	Medium-High (predictable timing)
Liquidation	1-5%	Extreme (can be catastrophic)

---

Feature Engineering

Six feature groups, 26 features total. Each group has a struct + compute function:

Group	Fields	Key Signals
Size (5)	size_zscore, size_quantile, size_vs_depth, size_vs_typical, size_raw	Large relative to depth = likely informed
Timing (4)	time_since_last_trade_ms, trades_in_last_1s, trade_rate_zscore, is_burst	Bursts (>10/sec) cluster with toxic flow
Aggression (4)	is_aggressor, crossed_spread_bps, depth_consumed_pct, sweeping_multiple_levels	Sweeping multiple levels = strongly informed
Flow (4)	signed_volume_imbalance_1s/10s, flow_autocorrelation, trade_aligns_with_flow	Flow autocorrelation detects persistent direction
Hyperliquid (5)	funding_rate, oi_change_1m/5m, trade_against_funding, near_liquidation_level, time_to_settlement_s	Trade against funding often = informed
Cross-Exchange (4)	binance_hl_spread_bps, binance_price_change_100ms/500ms, binance_leading	Binance leading >2bps = arb flow incoming

TradeFeatures combines all groups and provides to_vector() -> Vec<f64> with appropriate scaling (log for time, bps for funding, cyclical sin for settlement time).

See implementation.md for all struct definitions and compute functions.

---

Classifier Architecture

Option 1: Logistic Regression (Interpretable)

P(informed) = sigmoid(w . x + b) -- simple dot product of weights with feature vector. Good starting point for interpretability.

Option 2: Small MLP (More Expressive)

2-hidden-layer MLP (input -> 32 -> 16 -> 4 classes) with ReLU activations and softmax output. predict_informed_prob() returns P(informed) + P(liquidation) since both are toxic to market makers.

See implementation.md for both classifier implementations.

---

Training

Standard mini-batch training with Adam optimizer (lr=0.001), batch size 64, 100 epochs with early stopping on validation AUC. Evaluation uses binary toxic vs non-toxic classification (Informed + Liquidation = toxic).

See implementation.md for training loop and AUC computation code.

---

Real-Time Integration

AdverseSelectionAdjuster wraps the classifier for real-time use:

• on_trade(): Classifies each trade and updates an EMA of informed probability (alpha ~0.1, decays over ~10 trades)
• get_kappa_adjustment(): Reduces kappa proportional to informed flow intensity (floors at 0.3 -- never reduce by more than 70%)
• get_spread_adjustment_bps(): Widens spread proportional to informed intensity
• should_go_defensive(): Returns true when informed intensity exceeds high_toxicity_threshold

See implementation.md for full AdverseSelectionAdjuster code.

---

Liquidation Detector (Specialized Subsystem)

Liquidations deserve their own detector because they're:

• Highly predictable from OI + funding
• Extremely toxic
• Need fast response (pull quotes, don't just widen)

LiquidationDetector tracks OI and funding history in ring buffers and computes cascade probability from four additive signals:

Signal	Contribution	Typical Threshold
Rapid OI drop (1m)	+0.3	OI change < -2%
Rapid OI drop (5m)	+0.3	OI change < -5%
Extreme funding	+0.2	Funding percentile > 95th or < 5th
Funding-direction squeeze	+0.2	Extreme funding + OI drop > 1%

Decision thresholds: is_cascade_active() at prob > 0.5, should_pull_quotes() at prob > 0.7. Max probability capped at 0.95.

See implementation.md for full LiquidationDetector code.

---

Validation

Key Metrics

Track ClassifierValidation with: AUC-ROC, precision@50% recall, per-class confusion matrix/precision/recall, Brier score, information ratio, and economic metrics (PnL with/without classifier, improvement %).

See implementation.md for the ClassifierValidation struct definition.

Acceptance Criteria

• AUC > 0.65 (informed vs noise classification)
• Information Ratio > 1.0 (adds value vs base rate)
• PnL improvement > 10% vs no classifier

---

Dependencies

• Requires: measurement-infrastructure, signal-audit
• Enables: Dynamic kappa adjustment, spread widening on toxic flow

Common Mistakes

1. Overfitting to size: Large trades aren't always informed
2. Ignoring timing: Clusters of trades are more toxic than isolated ones
3. Missing cross-exchange: Arbitrage flow is predictable from Binance
4. Static thresholds: Liquidation conditions vary with market state
5. Binary thinking: Use probabilities, not hard classifications

Next Steps

1. Build labeled dataset from historical fills + price evolution
2. Run signal audit to identify predictive features
3. Train classifier (start with logistic regression for interpretability)
4. Validate AUC > 0.65, IR > 1.0
5. Integrate into quote engine
6. Build liquidation detector as separate fast path
7. Monitor classifier decay (retrain monthly)

Supporting Files

• implementation.md -- All Rust code: feature structs, classifiers, training loop, real-time integration, liquidation detector, validation

---

Known Issues from Production

• Feb 12 (Mainnet): 2x AS overestimate — AS classifier double-counted in certain flow conditions, leading to spreads wider than necessary. The fix introduced markout-based AS exclusively (deferred fill-time AS).
• Feb 18: Binary side-clearing via TOXICITY CANCEL — Code path cleared ALL quotes on one side when toxicity was high. This created a death spiral (no bids → position can't reduce → more toxicity → permanent quote loss). TOXICITY CANCEL was deleted. Route toxicity through AS spread multipliers instead. Binary side-clearing is ALWAYS wrong.
• Feb 20: AS deduplication — Vol compensation 0.5*gamma*sigma^2*T is already in GLFT. Must subtract it from measured AS before adding the residual: as_net = max(0, as_raw - vol_floor). Without dedup, AS and GLFT vol compensation double-count.