Kit Catalog

Adaptive boosting of weak learners weighted by residual magnitude.

Arm a real alert row from a Predicate-shaped condition. Browser execution never writes; server execution persists through the injected authenticated alerts writer.

Against-the-spread cover rate: wins / (wins + losses).

Mean of the edge column across all applied bets.

Backfill a source over a date range. Browser preview processes the first 100 rows and reports the total; server processes the full range.

Kelly scaled by current bankroll relative to a target: shrink stakes during drawdown, expand during streak.

Bayesian neural network approximation via Monte Carlo Dropout. Yields both a point estimate and a calibrated uncertainty — high-variance predictions can be sized down or skipped. Toy task: noisy sine regression.

Weight each base model by posterior probability proportional to exp(−loss).

Run BayesianRidge on engineered features and emit posterior-mean predictions as a derived signal column.

Projects how fast an opening edge erodes as the line moves, so a stale edge is never bet as a live one.

Calibrates probabilities via a beta-distribution warp; often generalizes better than Platt for skewed score distributions.

Browser-safe adapter from BetSlate to Frame.

Browser-safe adapter from BetSlate to Metric.

Browser-safe adapter from BetSlate to Predictions.

Browser-safe adapter from BetSlate to Series.

Bayesian NN via MC dropout: variance from N stochastic forward passes gives a calibrated CI.

Browser-only pure-JS Bradley-Terry rater fit by Minorization-Maximization. Estimates latent multiplicative strengths so that P(i beats j) = π_i / (π_i + π_j).

Mean squared error between predicted probability and binary outcome.

Cache the inbound body by key/content-hash with a TTL. Browser uses an in-memory map; server reads/writes the run_cache store via ctx.cache.

Browser-safe adapter from CalibrationParams to Frame.

Browser-safe adapter from CalibrationParams to Series.

Ordered-boosting classifier built for categorical features; strong out-of-the-box without tuning.

CLV in American-odds cents: priceAtPick − priceAtClose.

CLV in points: (closingLine - pickLine) × side. Proxy for true edge.

Route a pipeline by a predicate. Evaluates once and takes exactly one branch (the other is never evaluated). Behaves identically in browser and server.

Stake scales with prediction confidence: stake = bankroll × scale × |p − 0.5|. No-edge-aware.

Each model contributes proportional to |p − 0.5|. Confident models dominate the blend, uncertain ones drop out.

Shift predicted probability by the residual-quantile bias measured on calibration set. Distribution-free.

Empirical residual quantile for 90% coverage prediction interval.

Predicts the training-set mean (or a configured constant) for every row. Sanity check: anything that beats this is doing real work.

Drops features whose pairwise |correlation| exceeds threshold (default 0.95).

Computes binary cover label and spread-derived signals (f_spread_line, f_is_big_favorite, f_is_pick).

Browser-only axis-aligned binary decision tree fit via greedy Gini-impurity splits, max-depth + min-samples stopping. A single interpretable tree.

Reduce weight on highly-correlated base models: w_i ∝ 1 / (1 + mean correlation with others).

Monitor feature drift for a model. Browser preview measures drift on a 200-row sample; server measures over the full history. Flags when drift exceeds the threshold.

Correlation between edge in window t and edge in window t+1. Measures whether edges decay over time.

Stake proportional to edge: stake = bankroll × scale × max(edge, 0).

Linear model with L1+L2 penalty; balances Lasso and Ridge for sparse, stable fits.

Adjusts a raw Elo differential for rest, travel, QB-out, and an optional explicitly enabled team-HFA estimate, then derives a predicted scoring margin and home win probability.

Maintains team Elo ratings and exposes home/away rating pair as features.

Mean per-bin gap between predicted probability and observed frequency, weighted by bin size.

Compute an exponentially weighted moving average of a numeric column, optionally partitioned by a group key and ordered by a sortable key. Pure, leakage-safe; the EMA at row t uses only rows ≤ t within its group.

Ensemble of extremely randomized trees; cheaper variance reduction than Random Forest.

Edge-independent flat stake (default 1% of bankroll).

All-or-nothing: bet a fixed unit only when edge ≥ threshold; otherwise skip. Simple, transparent.

0.25× Kelly stake by default; lower variance than full Kelly.

Browser-safe adapter from Frame to BetSlate.

Browser-safe adapter from Frame to Metric.

Browser-safe adapter from Frame to Predictions.

Browser-safe adapter from Frame to Series.

Full Kelly stake: bankroll × (b·p - q)/b. Maximizes log growth; high variance.

TF.js-compatible tabular feature transformer for browser-side use.

Graph Attention Network for team matchups. Each game is a 2-node graph (home, away); multi-head attention learns pairwise interaction weights between team features. Toy task: predict home margin from two team feature vectors with an asymmetric interaction term.

Allows a model onto Tonight only when it beats the fair-market ATS baseline on the same graded rows, has strong sample sufficiency, and every populated calibration bin is within its model-implied 95% binomial interval.

exp(mean(log(p))). Pulls down disagreement. Requires p in (0,1]; robust when one model is wildly wrong.

Graph neural network over a team-vs-team graph; captures schedule-strength + transitive matchup effects.

sklearn GradientBoostingRegressor: sequential weak learners minimizing residuals; robust default.

Browser-only gated recurrent unit for sequence regression. Lighter and faster than LSTM with comparable accuracy on short-window time series.

0.5× Kelly stake. Roughly halves variance vs full Kelly for ~75% of the growth.

Binned calibration: each predicted probability is replaced with the empirical outcome rate in its bin.

Cover% within high-confidence bucket only (default p > 0.6 or p < 0.4). Tells you whether confident picks are reliable.

Separates a stat into home-only and away-only rolling averages (capturing venue-dependent skill).

Measures how far a pick vector sits from the house consensus, so an uncrowded edge is not priced like a copied one.

Pass-through. Use when downstream consumers must always run a calibration block.

Non-parametric monotone calibration via pool-adjacent-violators.

ROI computed against Kelly-sized stakes (not flat units). Reflects compounding when staking varies.

Browser-only k-nearest-neighbors classifier. Stores the training set, predicts via majority vote of the k closest rows. Useful when the decision surface is local and non-linear.

LassoCV-based feature selection on 70% train; keeps features with |coef| > threshold.

Histogram-based gradient boosting; fastest of the boosters on large tabular data with categorical features.

Ordinary least squares; transparent baseline for spread/total margin prediction.

OLS-fit predictions of a target column from a feature subset; appended as f_lr_predicted.

-mean(y·log(p) + (1-y)·log(1-p)). Penalizes overconfident wrong predictions.

Linear classifier with L2 regularization; baseline for binary outcomes.

Maximum consecutive wins in the result series.

Run an explicit bounded parameter search against a caller-supplied objective scorer, returning the best real scored trial.

Longest run of consecutive losing picks, ordered by input position. Useful for stress-testing bankroll resilience and detecting bad-luck clusters.

Stateful sequence model trained client-side via TF.js. Toy task: next-step sine-wave prediction. Production task: PBP / drive-level sequence regression.

Blends per-book de-vigged win probabilities into a weighted-mean consensus probability plus a dispersion measure of market agreement.

Spread vs. consensus arbitrage signals and line movement features.

Browser-only pure-JS Massey rating system. Solves the linear least-squares system MᵀMr = Mᵀb via Cholesky factorization. Output: per-team additive rating recovered from observed game margins.

Largest peak-to-trough drop in bankroll. Wraps analytics2/metrics/formulas.drawdown.

Worst-bin gap between predicted probability and empirical accuracy across N bins.

Per-row median of base model predictions. Outlier-resistant; non-differentiable.

Browser-safe adapter from Metric to BetSlate.

Browser-safe adapter from Metric to Frame.

Browser-safe adapter from Metric to Predictions.

Browser-safe adapter from Metric to Series.

Scale a column to [0, 1]: (x − min) / (max − min). Adds <col>_mm to the frame.

Sparse feature extraction for low-data regimes; defaults safe in cold start.

Browser-only MDN. Predicts a 3-component Gaussian mixture (π, μ, σ) per row via custom log-sum-exp NLL loss. Handles bimodal / boom-or-bust outcomes that a single-Gaussian regressor must average over.

MLB team-level rolling runs, OPS, ERA, ballpark factor; schedule-aligned for game prediction.

MLB PA/AB, LHP/RHP splits, ballpark factor, 10-game rolling stats.

Browser-runnable multi-layer perceptron via TensorFlow.js.

Browser-safe adapter from ModelArtifact to BetSlate.

Browser-safe adapter from ModelArtifact to Frame.

Browser-safe adapter from ModelArtifact to Metric.

Browser-safe adapter from ModelArtifact to Predictions.

Browser-only multi-task head with a shared dense trunk and three specialised heads (cover / total / moneyline) trained jointly via weighted-sum loss. Exploits target correlation when the same game produces all three lines.

Browser-only Gaussian Naive Bayes. Estimates per-class mean and variance per feature, predicts argmax of log P(x|c) + log P(c). Cheap and fast when features are approximately Gaussian within each class.

NBA team-level rolling stats, pace/eff splits, schedule-alignment features (game-level, not player).

NBA per-player minutes, scoring avgs, team pace, position norms.

NFL scouting combine results (season, draft_year, draft_team/round/ovr, player_name, pos, school, ht, wt, forty, bench, vertical, broad_jump, cone, shuttle).

Player contract terms (player, position, team, is_active, year_signed, years, value, apy, guaranteed, apy_cap_pct, draft background).

Weekly team depth charts (season, week, club_code, depth_team, full_name, position, depth_position, gsis_id, jersey_number, formation).

Draft-pick trade-value curves by pick number, multiple published models (stuart, johnson, hill, otc, pff).

Draft-pick outcomes (season, round, pick, team, player, position, college, career AV, games, honors) for prospect-pipeline modeling.

Per-team NFL home-field-advantage estimates from recency-weighted home/away closing-line error, shrunk toward the league mean. Held-out validation did not beat flat 48 Elo, so this feature is opt-in only.

Play-level manual charting data (formation, motion, play-action, RPO, screen, pressure/scheme flags) keyed to nflverse play/game ids.

Officiating crew assignments per game (game_id, official position, official name, season).

Live NFL weekly injury reports (nfl_injuries, 90k+ rows, nflverse-sourced). No params: current week, active statuses only. season+week params: that historical week, for training-harness use across many weeks. Point-in-time caveat: nflverse dropped date_modified for 2025+ rows -- report recency is provable only via our own updated_at (pipeline-alive), never the report's own timestamp; see nfl-injuries.ts for the full honesty split.

Live NextGen Stats (tracking-derived advanced metrics) across passing/receiving/rushing, selected by the required statType param.

Live Pro-Football-Reference advanced stats (pressure/blitz/hit rates, yards before/after contact, broken tackles, drop rates) across passing/rushing/receiving x season/week, selected by required statType + period params.

Live NFL play-by-play (curated column subset: down/distance/field position/play type/EPA/WP/success), requires season or gameId to avoid an unbounded scan of a 1.24M-row table.

Per-player rolling stats, opponent strength, snap counts.

Cross-source player id crosswalk (gsis/pfr/espn/sleeper/yahoo/otc/pff/etc ids) plus basic bio -- join key for combining across nflverse tables.

Master player reference table (identity, bio, position, draft background, career span, latest team, status).

ESPN Total QBR per QB per week/season (qbr_total, pts_added, epa_total, qb_plays, rank, qualified).

Per-team rest-days-since-last-game, reshaped from nfl_schedules.home_rest/away_rest into one row per team per game (joins on season+week+team like every other team-week data Block). is_short_week (<=6 days) and is_extra_rest (>=13 days, i.e. off a bye) are labeled convenience flags, not a modeling claim.

Season-total player box-score + efficiency stats (passing/rushing/receiving volume and advanced metrics, fantasy points).

Season-level roster membership (team, position, status, bio, draft background) per player-season.

Per-player, per-game offense/defense/special-teams snap counts and percentages.

Provenanced NFL venue coordinate reference data (30 rows, all 32 current franchises; Wikidata-sourced, CC0). Ignores its input and emits the registry as a Frame -- a data source, not a transform. Backs deriveTravelDistance (src/lib/lenses/travel-distance.ts).

Static team reference (abbreviation, full name, nickname, conference, division, colors, logo urls) -- 32-ish rows, no filters needed.

Per-game player box-score stats (passing/rushing/receiving volume and advanced metrics, fantasy points) -- the weekly counterpart to nfl-seasonal-stats.

Week-level roster membership (team, position, status) -- the weekly counterpart to nfl-seasonal-roster, largest roster table (900k+ rows).

Walk-forward-safe isotonic recalibration followed by a nearest-neighbor empirical NFL margin distribution, cover/push probabilities, and two-sided spread EV.

NHL TOI, PP/PK minutes, line matchups, goalie pairing.

De-vigs a two-way American-odds market via multiplicative normalization and returns both sides' fair win probabilities plus the vig removed.

Send a notification. Browser preview fires a UI toast only; server delivers via the real channel (Slack/email/toast).

Expand a categorical column into k binary indicator columns named <col>__<level>.

Subtracts opponent baseline from raw stat: f_adj = f - mean(f | opponent). Surfaces matchup-specific edge.

Play-by-play windowing for attention/transformer model inputs.

Browser-only pure-JS Plackett-Luce rater for full rankings. Fit by Hunter (2004) MM updates. Output: per-item log-strengths that imply the probability of any rank order — useful for DFS lineups and tournament finish predictions.

Two-parameter logistic-regression calibration; classic Platt 1999.

Browser-only generalized linear model with log link and custom Poisson NLL loss. λ = exp(W·x + b). Built for non-negative count outcomes like team goals or play counts.

Browser-safe adapter from Predictions to BetSlate.

Browser-safe adapter from Predictions to Frame.

Browser-safe adapter from Predictions to Metric.

Browser-safe adapter from Predictions to Series.

Gross wins / gross losses. >1 is profitable; >2 is rare and good.

Facebook Prophet time-series model: trend + weekly seasonality. Use for game-tempo or scoring drift.

Browser-only quantile regression with three heads (Q10 / Q50 / Q90) trained via pinball loss. Outputs full per-row interval — predicted lower / median / upper.

Ensemble of decision trees with bagging; robust default for tabular data.

Combines multiple model outputs by averaging per-row ranks; robust to scale differences across models.

Controls the false-discovery rate across many segment tests so a fluke pattern is never sold as an edge.

Days between consecutive games for each team. Adds f_home_rest, f_away_rest, f_rest_advantage (home − away).

Retrain a model on a dataset. Browser preview reports what would be retrained and on how many rows; server enqueues the job on the existing nflbets-training app.

L2-penalized linear regression; stable on collinear features.

Standard sportsbook ROI: (wins × 0.909 - losses) / (wins + losses + pushes).

Computes a rolling average over the last N rows for a given column.

Schedule a target job on a cron/interval. Browser preview fires the target once now and shows the next-run stub; server registers the recurring trigger.

Browser-safe adapter from Series to Frame.

Browser-safe adapter from Series to Metric.

Browser-safe adapter from Series to Predictions.

Permutation-invariant set encoder. Element-wise embed + self-attention + mean-pool. Toy task: sum-of-set binary classification on N=10 reals.

Mean return / std-dev of returns. Higher = more reward per unit volatility.

Empirical-Bayes shrinkage: stabilizes small-sample estimates by blending toward the global mean with weight 1/(n+k).

Reshape predicted probabilities through a temperature-controlled sigmoid: pulls extremes toward 0.5 or pushes them out.

Cross-sport ETL: rolling avgs, exponential moving avgs, schedule alignment.

Meta-model trained on base model predictions; learns optimal blend per row.

TabNet applies sequential soft attention to select which input features are most relevant at each step. Interpretable attention masks reveal which features the model used.

Single-parameter calibration: divides logits by T. T>1 softens overconfidence; T<1 sharpens.

Browser-only causal dilated conv1d stack (dilations 1·2·4). Receptive field grows exponentially, so the model sees lag-k dependencies an RNN of equal depth would miss.

Keep only the top K base models by training-set metric; average their predictions. Drops weak voters.

Encoder-only transformer for PBP / drive-level sequences. Stacks multi-head self-attention + FFN with residual connections and layer normalization, mean-pools the sequence.

Fractional Kelly with hard cap at maxBetPct (default 5%) of bankroll.

Open-to-close spread/total delta and implied-probability shift. Positive delta = home line moved toward favorite.

POST a JSON payload to a URL. Browser preview echoes the payload without sending; server POSTs for real (https-only, SSRF-guarded).

Linear combination of model outputs with explicit weights, normalized by sum |w|.

Turns a model probability and a two-way price into the residual against the de-vigged fair price, so the model learns only what the market has not already absorbed.

Measures each module by removing it and re-scoring the same picks, reporting the delta with a confidence interval instead of an importance score.

Gradient-boosted trees; competitive baseline for sports betting tabular data.

Standardize a numeric column: (x − mean) / std. Adds <col>_z to the frame.