add-signal

name	add-signal
description	End-to-end workflow for adding a new predictive signal to the market maker
disable-model-invocation	true
context	fork
agent	general-purpose
argument-hint	[signal-name] [prediction-target]
allowed-tools	Read, Grep, Glob, Edit, Write, Bash

Add Signal Workflow

Step-by-step process for adding a new predictive signal. Follow measurement-before-modeling: define what you're predicting, log it, measure baseline, then build.

Step 1: Define the Signal

Before writing code, answer:

1. What does it predict? (price direction, fill probability, adverse selection, volatility)
2. At what horizon? (100ms, 1s, 10s, 60s)
3. What is the prediction target metric? (Brier Score for binary, IR for continuous)
4. What is the baseline? (naive predictor, e.g., always predict 50% for binary)

Step 2: Implement the Estimator

Create a new file in src/market_maker/estimator/:

// src/market_maker/estimator/my_signal.rs

pub struct MySignalEstimator {
    // State for online estimation
    // Use EWMA, rolling windows, or Bayesian updates
}

impl MySignalEstimator {
    pub fn new(config: MySignalConfig) -> Self { ... }

    /// Update with new market data. Called on every relevant event.
    pub fn update(&mut self, data: &MarketData) { ... }

    /// Get current signal value, normalized to a useful range.
    /// Document units in the return type or name.
    pub fn signal_value(&self) -> f64 { ... }

    /// Confidence in the signal (0.0 = no data, 1.0 = fully warmed up)
    pub fn confidence(&self) -> f64 { ... }
}

Conventions

• Use _bps suffix for basis point values
• Use _s suffix for seconds
• Use EWMA with configurable halflife for online stats
• Add #[serde(default)] to any checkpoint fields for backward compat
• Include a warmup_observations: usize counter

Step 3: Register in mod.rs

Add the module to src/market_maker/estimator/mod.rs:

pub mod my_signal;
pub use my_signal::MySignalEstimator;

Step 4: Wire into Signal Integration

src/market_maker/strategy/signal_integration.rs is the central hub. Add your signal:

1. Add field to the struct that holds estimators
2. Call update() in the appropriate update method
3. Add signal output to compute_signals() or equivalent
4. Include in prediction logging

IMPORTANT: Only the strategy teammate edits signal_integration.rs in team mode. Others propose changes via messages.

Step 5: Add Model Gating

In src/market_maker/calibration/model_gating.rs:

1. Add your signal to the gating system
2. Set initial weight to 0.0 (disabled until validated)
3. Define MI threshold for activation
4. Signal should only contribute when should_use_model() returns true

Step 6: Add Prediction Logging

Ensure your signal's predictions are logged for calibration:

1. Add to PredictionRecord in checkpoint/prediction system
2. Log both the signal value and the prediction target outcome
3. Include market state conditioning variables

Step 7: Set Up Calibration

In src/market_maker/calibration/:

1. Add Brier Score tracking for your signal's predictions
2. Add Information Ratio computation
3. Set up conditional calibration (by regime, volatility, time of day)
4. Define warning thresholds (IR < 1.0 = warning, IR < 0.8 = critical)

Step 8: Verify

# Must pass
cargo clippy -- -D warnings
cargo test --lib

# Check signal compiles and integrates
cargo build

Step 9: Validate Live

1. Run paper trader with signal enabled (weight = 0.0 initially)
2. Collect at least 24h of prediction data
3. Compute Brier Score and IR
4. If IR > 1.0, gradually increase weight via model gating
5. Monitor for 1 week before full deployment

Checklist

• Signal has clear prediction target and horizon
• Estimator implements online update (no batch processing)
• Units documented in variable names
• Warmup tracking with confidence metric
• Registered in estimator/mod.rs
• Wired into signal_integration.rs
• Model gating configured (starts at weight 0.0)
• Prediction logging enabled
• Calibration metrics defined (Brier, IR, conditional)
• Checkpoint fields use #[serde(default)]
• Clippy clean, tests pass