Credit scorecards with scorecraft::CHEAT SHEET
The pipeline every stage is a function
Each stage returns an object that prints; every decision goes to a ledger. Two shortcuts chain the stages.
- 0
scr_split()type columns, split the sample - 1
scr_triage()profile, sentinels, early failures - 2
scr_bin()optimal bins and three gates - 3
scr_model()model votes and consensus - 4
scr_scorecard()logistic fit on WOE, points - 5
scr_align()raw score to the declared scale - 6
scr_cutoff()cut-off, strategy, rejects - 7
scr_sql()scoring in R and in SQL
scr_select() = 0 to 3scr_scorecard() = 4 and 5Configure
scr_config(preset = "moderate", ...) Every knob of every stage in one object. The preset sets how tight the funnel is; override any key by name.
cfg <- scr_config("moderate", nthread = 4)
| preset | variables | min votes | corr cut | IV floor |
|---|---|---|---|---|
aggressive | 10 to 15 | 3 | 0.6 | 0.03 |
moderate | 10 to 25 | 2 | 0.7 | 0.02 |
lazy | 10 to 40 | 1 | 0.8 | 0.02 |
scr_config_keys(stage = NULL) Every key with its stage (0 to 12), default and meaning.
scr_presets() This table. scr_verbose(on) Progress messages on or off.
Reading conventions
target 1 is bad; more points are safer (
higher_is_safer, odds safe:event).target 1 is good; more points, more likely (
higher_is_riskier, event:safe).objective never changes the selection; event_level decides which target value is the event.
Points per bin: sc$points
| vl_score_01 | WOE | points | |
|---|---|---|---|
(-Inf; 33.36] | -2.06 | 61 | |
(33.36; 38.15] | -0.73 | 22 | |
(38.15; 44.24] | -0.66 | 20 | |
(44.24; 48.06] | -0.52 | 16 | |
(48.06; 63.94] | 0.04 | -1 | |
(63.94; 72.61] | 0.70 | -21 | |
(72.61; +Inf] | 1.00 | -30 |
On scr_demo: score = 538 + points of 12 variables.
Select variables stages 0 to 3
res <- scr_select(scr_demo, "default", cfg, drop = c("id", "churn"), date_col = "ref_date")
scr_select(data, target, config, drop, date_col, event_level, export) Split, triage, binning and consensus in one call. With date_col the hold-out is out of time; without it, a stratified random 30%.
No candidate leaves the report: each one keeps the stage it failed at, and why.
Stage by stage
scr_split(data, target, date_col = NULL, ratio = 0.3) Type the columns and split train and hold-out.
scr_triage(split, config) Profile on train only. Fails CONSTANT, NEAR_CONSTANT, TOO_MANY_MISSING, HIGH_CARDINALITY, NO_SIGNAL, DUPLICATE_OF. A sentinel (-999) with mass and signal becomes a flag column x__sp.
scr_bin(triage, config) Optimal bins on train, in parallel by column, then three gates: the eight admission rules below; hold-out revalidation with frozen bins (IV ratio, PSI); redundancy pruning by rank correlation on the WOE space.
scr_model(bins, config) glmnet, xgboost, lightgbm and ranger vote; the consensus is weighted by each model's hold-out Gini and the shortlist stays in [target_min, target_max].
Read the result
scr_selected(res, which = "final") The shortlist; also "consensus", "manual".
scr_funnel(res, only_selected = FALSE) Every input column, its IV, KS, PSI and the reason it stopped.
scr_gains(res) Bin-level gains of the approved variables.
scr_leakage(res, threshold = NULL) Suspicious IV and degenerate bins.
scr_score_metrics(sc) AUC, KS and Gini per sample, with bootstrap CI.
scr_score_gains(sc, sample = "holdout") Gains per score band frozen on train: KS, lift, pct_event, pct_nonevent and woe = ln(pct_event / pct_nonevent), > 0 when the band rate is above the overall rate; odds in the scale orientation.
summary(res), plot(res), as.data.frame(res) Executive summary, funnel bar chart, funnel table.
Build the scorecard stages 4 and 5
sc <- scr_scorecard(res, base_score = 600, base_odds = 50, pdo = 20, challenger = "xgboost") sc$alignment # the fitted scale map
scr_scorecard(x, features, base_score, base_odds, pdo, direction, challenger, points_style) Logistic regression on the WOE columns with a sign check (a non-positive coefficient leaves, one at a time), points per bin, bootstrap CI, bands frozen on train, PSI and CSI.
factor = pdo / ln 2
offset = base_score − factor·ln(base_odds)
scr_align(raw, y, base_score, base_odds, pdo, direction, method = "regression") Align the raw score of any engine: empirical log-odds regressed on score bands, composed with the PDO map. Two scorecards aligned this way compare point for point.
predict(align, raw, type = "score") Points, or type = "prob" for the implied probability.
Challenger ("xgboost", "lightgbm"): aligned to the same scale for comparison, with supports_scorecard = FALSE: no points, no reason codes. Points style: "base_plus_deviation" or "distributed".
Decide stage 6
scr_cutoff(sc, n_cuts = NULL, cuts = NULL) Approval, event rate on each side, events avoided and KS at each cut. Cuts are train quantiles applied frozen to the hold-out.
scr_strategy(sc, revenue_good = 1080, loss_bad = 4500) Bands with volume, event rate, odds_event, log_odds, decision and expected profit per account:
p* = revenue_good / (revenue_good + loss_bad)
scr_strategy(sc, rule = "crossing") Cut where the event and non-event distributions cross (max KS, st$crossing). Propensity: most likely band first; target, review, skip.
scr_reject(sc, population = NULL, accepted = NULL) Honest reject inference: population scope, outcome coverage per band and a sensitivity band (the rejects 2, 4 or 8 times worse), never a single invented multiplier.
Coarse classing lab manual bins
Manual bins and manual variable choice, each with a reason, benchmarked against the optimal bins on train and hold-out.
Verdict: ACCEPTABLE, REVIEW or BLOCKED. The reason is mandatory and the ledger is append-only.
lab <- scr_coarse_classing(res) p <- scr_classing_propose(lab, "ds_region", groups = list(edge = c("NORTH", "SOUTH"), core = c("EAST", "WEST", "CENTRE"))) lab <- scr_classing_accept(lab, p, reason = "edge/core is what pricing uses") lab <- scr_classing_choose(lab, drop = "vl_score_10", reason = "not available at decision time") res2 <- scr_classing_apply(lab) # scr_result sc2 <- scr_scorecard(res2)
scr_coarse_classing(x, features = NULL, max_iv_loss = NULL) Open the lab on every variable that reached binning.
scr_classing_view(lab, variable) Current bins, train and hold-out IV.
scr_classing_propose(lab, variable, ...) One instruction per call:
scr_classing_accept(lab, proposal, reason, override = FALSE) and scr_classing_discard(lab, proposal, reason) Record the decision; override accepts a BLOCKED one.
scr_classing_choose(lab, keep, drop, force, reason) The final variable list by hand.
scr_classing_apply(lab) Commit; the scorecard, the R scoring and the SQL follow unchanged.
scr_classing_spec(lab, file), scr_classing_read(file), scr_classing_import(lab, file) Specification round trip through CSV or xlsx, for a business reviewer.
scr_decisions(x) The decision ledger of a lab, result or scorecard.
Metrics on any vector
scr_metrics(score, y, higher_is_event = TRUE, n_boot = 200) AUC, KS and Gini with a bootstrap CI.
scr_psi(base, compare, n_groups = 10, alpha = 0.05) PSI with the fixed thresholds and the sample-size-adjusted critical value.
scr_iv(g, y, laplace = 0.5) Information value of any grouping.