ML Inference Strategy

InferenceStrategy generates buy/sell signals from a 12-feature order-book vector using either a pre-trained ONNX model or an online Adam-SGD classifier that trains itself in real-time.

How it works

On every on_book call, FeatureComputer builds a 12-element feature vector from the current book snapshot and recent trade history.
RunningNormalizer standardises the vector to approximately zero mean / unit variance (active after a 30-tick warm-up; raw features are used before that).
The normalised vector is either:
- ONNX path — fed to a tract-onnx plan loaded from disk.
- Online fallback — classified by an incremental 3-class softmax trained with Adam via delayed self-labelling (label assigned after lookahead_ticks ticks using mid-price movement).
[sell_prob, hold_prob, buy_prob] drives a threshold gate (0.65 for ONNX, 0.60 for online); a signal is emitted only when one class exceeds the threshold.

Feature vector (12 features)

Index	Feature	Range
0	Order imbalance `(bid_qty − ask_qty) / total`	[−1, 1]
1	Spread bps / 100	[0, ∞)
2	RSI-14 / 100	[0, 1]
3	MACD histogram (sign × clamped magnitude)	[−1, 1]
4	Depth ratio top-5 bid / total	[0, 1]
5	EMA-50 deviation `(mid − ema) / ema`	uncapped
6	ATR-14 / mid	[0, ∞)
7	VWAP deviation, 100-trade rolling	uncapped
8	Depth slope `(bid[0] − bid[4]) / mid`	uncapped
9	Trade-flow imbalance, 20-trade rolling	[0, 1]
10	Kyle’s Lambda (tanh-squashed price-impact coeff)	(−1, 1)
11	VPIN trade toxicity	[0, 1]

compute_extended() appends a 13th feature (index 12): Hawkes trade-arrival intensity.

Online normalisation

RunningNormalizer runs Welford’s one-pass algorithm to accumulate per-feature mean and variance. Features are normalised in-place before reaching the classifier:

z_i = (x_i − μ_i) / max(σ_i, 1e-8)

Normalisation activates after 30 observations. Before warm-up, raw features are used unchanged.

Online learning (no ONNX model)

When no model file is supplied, the strategy trains an in-process softmax classifier with the Adam optimizer (β₁ = 0.9, β₂ = 0.999, ε = 1e-8, lr = 0.001, L2 = 1e-4):

Each tick the feature vector is stored alongside the current mid price.
After lookahead_ticks ticks, the mid-price change determines the label: +1 (price rose > threshold), -1 (fell), 0 (flat).
An Adam gradient step updates the 3-class weight matrix (3 × 12).
Signals are emitted only after min_updates_before_signal updates (default 25) to avoid noise during cold-start.

#![allow(unused)]
fn main() {
let config = OnlineLearningConfig {
    lookahead_ticks: 10,
    threshold: 0.0005,           // 0.05% price move to assign a label
    min_updates_before_signal: 25,
    classifier: OnlineClassifierConfig {
        learning_rate: 0.001,
        l2: 0.0001,
        beta1: 0.9,
        beta2: 0.999,
        epsilon: 1e-8,
    },
};
let strat = InferenceStrategy::new("BTCUSDT", dec!(0.01), None, Some(bus))
    .with_online_config(config);
}

Loading an ONNX model

cargo run --bin mercury -- run \
  --symbol BTCUSDT \
  --paper \
  --strategy inference \
  --model path/to/model.onnx

The expected ONNX input shape is [1, 12] with output shape [1, 3] (softmax over [SELL, HOLD, BUY]). A [1, 2, 20] LOB CNN model is also supported via InferenceStrategy::new_lob_cnn.

Model checkpointing

Online classifier weights can be saved and restored between sessions:

#![allow(unused)]
fn main() {
// Save after a live session
clf.save("btcusdt_weights.json")?;

// Restore at startup
let mut clf = OnlineClassifier::new(OnlineClassifierConfig::default());
clf.load("btcusdt_weights.json")?;
}

The checkpoint format is compact JSON containing updates, bias[3], and weights[36] (3 classes × 12 features).

Keyboard shortcuts