Model Cards¶

LightGBM Primary Classifier¶

Purpose: Predict direction of the next 5-min bar: short (−1), neutral (0), or long (+1).

Attribute	Value
Class	`tessera.models.lightgbm_model.LightGBMPrimary`
Framework	LightGBM 4.x
Objective	`multiclass` (3 classes)
Input	20 engineered features (see Features)
Output	Class probabilities \(P(\hat{y} \in \{-1, 0, 1\})\)
Training	Uniqueness-weighted samples (AFML §4)
CV	PurgedKFold, then CPCV 15-split for evaluation
HPO	Optuna TPE, 200 trials, maximise OOS Sharpe
Reproducibility	`seed_everything(42)`, tracked in ModelCard

Hyperparameters (best trial)¶

Parameter	Value	Notes
`n_estimators`	800	Early stopping on val set
`learning_rate`	0.03	Most important HPO parameter (fANOVA: 38%)
`num_leaves`	63	2nd most important (24%)
`min_child_samples`	40	Guards against overfitting on small folds
`subsample`	0.85	Row subsampling
`colsample_bytree`	0.80	Column subsampling
`reg_lambda`	1.0	L2 regularisation

Monotonic constraints¶

Domain knowledge allows positive monotone constraints on: - ofi_5m — higher OFI should not decrease long probability - universe_rank — higher rank should not decrease long probability

These constraints prevent the model from learning spurious sign reversals that survive in-sample but fail in regime shifts.

OOS performance (CPCV, 2021–2024)¶

Metric	Value
Annualised Sharpe	1.28
Deflated Sharpe	0.84
Win rate	52.1 %
Avg holding period	52 min

LightGBM Meta-Classifier¶

Purpose: Predict whether the primary model's direction call is correct (binary: correct / incorrect).

Attribute	Value
Class	`tessera.models.lightgbm_model.LightGBMMetaModel`
Objective	`binary`
Input	Primary model's OOS probability outputs + microstructure state
Output	Probability that the primary model is correct: \(p_{\text{meta}} \in [0, 1]\)
Training data	OOS predictions only — the meta-model is never trained on in-sample primary model outputs

Key design constraint: OOS-only training¶

The meta-model is trained exclusively on the primary model's out-of-sample predictions from the same purged folds.

Why this matters: Training on in-sample predictions would teach the meta-model the primary model's training biases rather than its generalisation behaviour, inflating meta-model precision by 8–12 percentage points (a bug found and fixed — see Pitfalls).

Role in sizing¶

The meta-model output \(p_{\text{meta}}\) feeds directly into the position size:

size = fractional_kelly(p_meta, win_loss_ratio, fraction=0.25) × vol_target_scale

When \(p_{\text{meta}} > 0.65\), the position is at full Kelly-adjusted size. When \(p_{\text{meta}} < 0.55\), the signal is not traded (meta-model veto).

OOS performance (CPCV)¶

Metric	Value
Ensemble Sharpe (primary + meta)	1.41
Improvement over primary alone	+0.13 Sharpe
Meta precision at threshold 0.60	61.4 %
Trades filtered (meta veto)	23 % of primary signals

Ensemble¶

Purpose: Combine primary and meta outputs into a final signal.

Attribute	Value
Class	`tessera.models.ensemble.EnsembleModel`
Method	Weighted combination, weights from CPCV Sharpe contribution
Formula	\(\hat{y} = w_p \cdot \hat{y}_{\text{primary}} + w_m \cdot p_{\text{meta}}\)
Default weights	\(w_p = 0.6,\ w_m = 0.4\)

The ensemble is the default deployed configuration.

PatchTST¶

Purpose: Sequence-model baseline for triple-barrier classification.

Attribute	Value
Class	`tessera.models.patchtst.PatchTSTClassifier`
Reference	Nie et al., ICLR 2023
Architecture	Patch transformer encoder + classification head
Parameters	≤ 5M
Input	Sliding windows of 60 bars × (all features)
Output	Class probabilities
Training	GPU-accelerated, AdamW, cosine schedule, 50 epochs

Architecture details¶

lookback = 60 bars
patch_len = 8 bars
→ 8 patches per feature sequence (padded to 64)

Encoder:
  d_model = 128
  n_heads = 4
  n_layers = 3
  ffn_dim = 256
  dropout = 0.1

Classification head:
  Linear(128 × n_patches, 3)
  Softmax → {-1, 0, +1}

Why PatchTST does not dominate LightGBM¶

Features already encode history: VPIN, realised vol, and UniverseRank summarise recent bar history into a single number. The transformer's 60-bar lookback adds redundant context.
Short effective lookback: Predictive information in 5-min crypto returns decays within 5–15 bars (per AFML §3 studies). A 60-bar window is mostly noise.
Larger HPO search space: PatchTST has batch size, learning rate, dropout, and patch length as additional hyperparameters, inflating the trial count and lowering the deflated Sharpe even when raw Sharpe is similar.

OOS performance (CPCV)¶

Metric	Value
Annualised Sharpe	1.32
Deflated Sharpe	0.76
Relative to LightGBM	−0.09 raw Sharpe, −0.11 DSR

Chronos Zero-Shot¶

Purpose: Foundation model baseline — no training, pure zero-shot signal extraction.

Attribute	Value
Class	`tessera.models.chronos_zeroshot.ChronosZeroShot`
Base model	`amazon/chronos-bolt-base` (T5, ~200M params)
Mode	Zero-shot — no fine-tuning on Tessera data
Reference	Ansari et al., TMLR 2024
Input	Log-return series only (univariate)
Output	Quantile forecast → converted to direction signal

Signal extraction¶

# Median quantile forecast → sign → direction signal
forecast = pipeline.predict(context=returns[-512:], prediction_length=1)
median = forecast[0].median(dim=0).values.item()
signal = +1 if median > threshold else (-1 if median < -threshold else 0)

Why Chronos underperforms¶

Reason	Detail
Pre-training distributional mismatch	Chronos was pre-trained on daily/weekly series (M4, ETT, electricity). 5-min crypto returns are near-white-noise at any horizon > a few bars.
Mean-reversion prior	Chronos implicitly learns that series revert to their local mean — correct for electricity demand, wrong for trending perpetuals
Strictly univariate	Cannot condition on VPIN, OFI, funding Z-scores — the features that drive the LightGBM edge
Zero-shot premise	Fine-tuning would partially bridge the gap but defeats the zero-shot premise and requires the same CV discipline as a supervised model

OOS performance (CPCV)¶

Metric	Value
Annualised Sharpe	0.72
Deflated Sharpe	0.51
Relative to LightGBM	−0.69 raw Sharpe

Conclusion: Chronos is not rejected categorically. If Tessera migrates to raw L2 order book data (tick-level), a convolutional sequence model trained end-to-end might be competitive. At the current tabular-feature abstraction level, the transformer adds complexity without adding signal.

Model Registry¶

Every promoted model is saved with a ModelCard (JSON) containing:

{
  "name": "lightgbm_primary",
  "version": "0.3.1",
  "type": "primary",
  "git_commit": "d576cb9",
  "training_date": "2026-05-17T14:30:00Z",
  "data_version": "2021-01-01:2024-12-31",
  "cv_scores": {
    "mean_sharpe": 1.28,
    "std_sharpe": 0.31,
    "deflated_sharpe": 0.84,
    "n_trials": 200
  },
  "hyperparameters": { ... },
  "feature_names": [ ... ]
}

Models are versioned under models/<name>/<run_id>/. A model is promoted to production only if its deflated Sharpe exceeds 0.75.