Canonical docs
In the research projects
The Python project computes next-bar return separately inside each split. The Compiled project fits a quantile target on training observations and reuses its cutoff. The target reference preserves each alternative’s parameters, source columns and forward-label boundary.
What this package owns
TradeOutcomeTarget consumes the shared execution contract through an injected TradeTargetContext and exposes private OutcomeLabels. See simulated trade outcomes for configuration, completion masks and feature isolation.
Owns train-fitted and transform-time target builders, including binary classification targets, return targets, breakout targets, barrier targets, and identity passthrough targets. Does not own feature engineering, model fitting, calibration, or execution strategy.
Key entry points
| Entry point | Use case | Notes |
|---|---|---|
QuantileBinaryTarget |
Labels from a train-fitted quantile cutoff | Fits the cutoff on train only. |
ThresholdBinaryTarget |
Labels from a fixed threshold | Supports explicit return thresholds. |
TradelineLongBinaryTarget |
Long-entry label from price-line context | Pairs with line-based feature helpers. |
NextReturnTarget, VolNormalizedReturnTarget, ForwardVolNormalizedReturnTarget |
Continuous forward-return style targets | Common for regression or ranking experiments. |
NextBarUpTarget, NextBarDownTarget |
One-bar direction labels | Classification baselines. |
ForwardBreakoutTarget, EmaBreakoutTarget |
Breakout-oriented labels | Uses forward path or EMA context. |
TripleBarrierTarget |
Path-aware barrier label | Requires enough rows for max_horizon and the volatility warmup. |
RiskRewardRatioTarget, ExitQualityTarget |
Outcome labels from upstream trade-path columns | Require the expected input columns to exist on every split. |
RandomBinaryTarget, IdentityTarget |
Controls or already-present target columns | Supports tests, baselines, and externally labelled data. |
Adjacent modules
limen.transformscan derive target inputs before target construction.limen.experiment.Manifestwires target builders intoprepare_data().limen.calibrationevaluates predictions against the target labels after model output.
Quick orientation
targets/├── README.md├── __init__.py├── _barrier_paths.py├── _perturbation.py├── ema_breakout.py├── exit_quality.py├── forward_breakout.py├── forward_vol_normalized_return.py├── identity.py├── next_bar_down.py├── next_bar_up.py├── next_return.py├── quantile_binary.py├── random_binary.py├── risk_reward_ratio.py├── threshold_binary.py├── trade_outcome.py├── tradeline_long_binary.py├── triple_barrier.py└── vol_normalized_return.pyThings to know
- Target classes receive the training split during construction and apply
transform()to every split. - Train-fitted targets must store only train-derived state before validation and test transformation.
- Forward-looking targets naturally create boundary nulls; manifest preparation drops null target rows downstream.
- Outcome-derived targets require their upstream columns to exist before target construction.