Financial ML

Leakage-free feature engineering, offline-train/online-infer deployment, triple barrier labeling, sample weighting, model staleness detection.

Skills: 38
Hand off: 289
Handed off from: 205

Hands off most often to

Brokers, venues and frameworks named here

Index

All 38 skills in this domain, alphabetically. Hover a row for its situation; open it for the full playbook.

A

adversarial-robustness-of-trading-signalsUse before promoting an ML trading signal, to measure how often epsilon-bounded feature perturbations flip its output. A black-box gradient-free test; for a differentiable net, true FGSM or PGD gives a tighter worst-case bound.trading alternative-data-feature-integrationUse when turning an alternative data source into model features and the event date differs from the date your fund actually received the data; enforces point-in-time lag mapping from knowledge date so the feature cannot see the future.alternative-data

B

backtesting-ml-models-against-transaction-costsUse when an ML model flips its prediction on nearly every bar and its Sharpe depends on that turnover; applies per-turnover cost charges and confidence thresholding before the strategy is judged. Order-level cost attribution is transaction-cost-analysis-tca-integration.tca

C

categorical-feature-encoding-for-instrument-identityUse when a cross-sectional model needs to know which instrument a row belongs to and the symbol column has hundreds or thousands of levels; smoothed target encoding computed out-of-fold, instead of one-hot columns that are almost all zero.categorical-encoding class-imbalance-handling-for-rare-signal-eventsUse when predicting rare events such as halts or flash crashes where one class is under a few percent; cost-sensitive weighting and undersampling with the probability recalibration that undersampling makes necessary.class-imbalance cold-start-handling-for-newly-listed-instrumentsUse when an instrument has too little price history to estimate risk from, such as an IPO, spin-off or new token, and a rolling volatility window would return NaN or a meaningless number that then sizes a position.cold-start concept-drift-vs-staleness-differentiationUse when a live signal degrades and you must decide what to fix, separating concept drift in the conditional relationship, covariate shift in the feature distribution, and plain data staleness, since each calls for a different remedy.covariate-shift cross-sectional-vs-time-series-model-designUse when choosing between a cross-sectional model that ranks instruments against peers at one timestamp and a time-series model that scores each instrument against its own history; the two imply different neutrality and sizing.cross-sectional

E

ensemble-signal-combination-without-overfittingUse when combining several alpha signals or sub-models into one score and unconstrained regression would produce extreme offsetting weights; causal normalisation, non-negative weights, 1/N shrinkage and inverse forecast-error-variance weighting.ensemble-learning explainability-for-live-trading-signalsUse when a live model score must later be explained to a risk officer or model reviewer; reconciles supplied feature attributions against the score the model actually emitted, ranks the drivers and writes a readable audit record.explainable-ai explainable-boosting-machines-for-regulated-signalsUse when a signal must be explainable to a model-risk reviewer; composes an already-fitted explainable boosting machine from its shape functions and audits declared versus realised interactions, so attribution survives review.ebm

F

feature-engineering-cost-benefit-trackingUse when a feature pipeline has accumulated expensive inputs with negligible marginal accuracy; weighs each feature's permutation importance against its licensing cost, compute cost and inference latency before pruning.feature-cost-benefit feature-engineering-without-leakageUse when designing or auditing features for a model that predicts a future price move, to guarantee each feature is computable strictly before the target is realised; catches same-bar returns and label-derived inputs.data-leakage feature-importance-drift-monitoringUse when a live model has a recorded training-time importance profile and you want early warning that regimes have reordered its drivers; measures what the model leans on, not whether it still makes money.feature-drift feature-selection-stability-across-foldsUse when a selector such as Lasso, Boruta or RFE runs inside cross-validation and you need to know whether its output is a property of the signal or of each fold; the chance-corrected Nogueira stability index.feature-selection feature-store-for-live-and-backtest-parityUse when a model trains offline on bars and serves online on a stream, and two separate feature code paths would drift apart; one shared computation engine plus a numeric parity check between batch and streaming output.feature-store

G

gradient-boosted-tree-vs-neural-net-tradeoffsUse when a signal's architecture is still open between gradient-boosted trees on tabular features and a deep network on raw sequences; produces an evidence-tagged prior from published benchmarks, to be tested rather than trusted.gradient-boosted-trees

H

hyperparameter-tuning-without-target-leakageUse when a search selects hyperparameters for a model whose labels are forward-looking and overlapping, such as h-bar returns or triple-barrier outcomes; purged and embargoed nested cross-validation so the selection folds stay clean.hyperparameter-tuning

L

label-noise-estimation-in-financial-targetsUse when a binary financial target such as a triple-barrier or fixed-horizon label is mislabelled by microstructure noise; confident learning estimates the noise transition matrix and flags suspect samples before training.confident-learning

M

model-card-documentation-for-trading-modelsUse when a trading model is promoted, revalidated or handed to someone who did not build it; a structured record of identity, intended and out-of-scope uses, training data, evaluation, known failure modes and owner.model-governance model-inference-latency-budget-for-live-tradingUse when deciding whether an ML model is fast enough to keep serving live signals; audits a captured latency sample against a percentile budget with nearest-rank P25 to P99.9 and a sample-count gate that refuses to answer on too few points.latency-budget model-monitoring-dashboard-for-non-technical-stakeholdersUse when a live model's health must reach a risk officer or portfolio manager who cannot interpret a PSI figure; grades accuracy, drift and staleness into plain statuses. Not a substitute for MiFID II RTS 6 real-time monitoring.monitoring-dashboard model-serving-infrastructure-ab-testingUse when a challenger model is evaluated against the champion in production; deterministic salted traffic routing, shadow execution and a Welch two-sample t-test, so promotion rests on a test rather than on luck.ab-testing model-staleness-detectionUse when an ML signal model is already live and needs continuous health monitoring: rolling realised accuracy with a confidence bound, binned PSI feature drift against the training baseline, and a sizing multiplier that decays as it degrades.model-monitoring model-training-data-freshness-slaUse as the gate immediately before a scheduled retraining job, when an upstream pipeline you do not control assembles the dataset; measures event-time ingestion lag against a target and breach ladder, netting out exchange non-publishing days.sla-monitoring model-versioning-and-rollbackUse when a model or execution algorithm is registered, promoted or rolled back in production; an append-only SHA-256 registry where one semantic version permanently identifies one artifact, plus a guarded rollback path.model-registry multi-horizon-forecasting-architectureUse when models forecast the same instrument over several forward horizons at once and their outputs must become one tradeable alpha; rescales each forecast onto a common horizon before weighting them.multi-horizon multi-model-ensemble-weight-decayUse when live capital is split across several forecasting models whose relative skill drifts with the regime; exponentially discounts each model's loss and information coefficient so a decayed model loses weight.ensemble

O

offline-train-online-infer-deploymentUse when a model trained in an offline pipeline must run inference inside a live bot process, to eliminate train-serve skew with a digest-verified artifact bundling weights, scalers and the feature contract.train-serve-skew online-learning-for-adaptive-signal-modelsUse when a live linear signal model must update observation by observation because the relationship moves faster than the retraining cadence; LMS, normalised LMS or recursive least squares with forgetting.online-learning

P

point-in-time-database-for-ml-training-dataUse when assembling a feature and label matrix, to as-of join features on the knowledge axis so no row carries a restated value that was not knowable at the label timestamp. The storage schema itself is backtest-database-schema-for-point-in-time-queries.point-in-time-db

Q

quantile-regression-for-uncertainty-aware-signalsUse when position sizing needs a signal that says how sure it is; predicts conditional return quantiles rather than a point forecast, so a confident and a guessing prediction are not sized identically.quantile-regression

R

regime-detection-for-strategy-switchingUse when several strategy variants are live and only one should be armed at a time; classifies each closed bar with Wilder ADX/DMI and an ATR volatility z-score, with hysteresis that also delays the risk-off switch.adx reinforcement-learning-safety-constraints-for-executionUse when a reinforcement-learning policy proposes execution quantities that reach an order router; shields them behind deterministic hard limits on order size, position cap, spread width and terminal state.safety-constraints reproducible-ml-training-pipelinesUse when a model will size, time or select trades and someone must later reproduce it exactly; records dataset, hyperparameters, code version and environment as SHA-256 digests and trains under scoped RNG seeding.mlops

S

sample-weighting-for-overlapping-labelsUse when labels span multiple bars and overlap, so consecutive observations share the same price moves; computes label concurrency and average uniqueness to weight samples. Fold-boundary leakage still needs purging.sample-weighting synthetic-labels-from-triple-barrier-methodUse when the supervised target must reflect the path rather than the endpoint; labels each event by which barrier it touches first, a volatility-scaled profit target, a stop, or a time limit.triple-barrier-method

T

transfer-learning-across-correlated-instrumentsUse when a new or thinly traded instrument has too little history to fit a model on its own and a liquid co-moving instrument exists; fits on the source and adapts, rather than fitting noise on the target.transfer-learning

Search the whole catalog