Sign in required

Same lift formula as the chart below, but the DNN side only counts transactions whose BIN appears in data/volaris_dnn_bin_range_allowlist.csv. The heuristic baseline is unchanged. Answers the question: "If DNN only served the BINs it was trained on, what would the lift look like?" Positive = DNN winning.

Daily transaction value lift of DNN over heuristic in USD. Positive = DNN winning. Lift = (DNN approval rate − Heuristic approval rate) × DNN volume × $360 avg txn (Volaris 90d avg).

Daily lift using real per-transaction USD amounts from VW_ATHENA_PAYMENT_ATTEMPT. Approval rates are $-weighted (sum approved $ / sum total $). Positive = DNN winning. Lift = actual $ DNN approved − (heuristic $-AR × DNN total $).

Non-DNN transactions sampled to match DNN volume per amount bin each day — removes transaction-mix confound vs raw heuristic comparison.

Of the transactions in the DNN experiment (experiment_id 767e0ebe), the share where the DNN model was actually evaluated — i.e. its ranked_processors + scores are saved in ATHIA_PREDICTIONS.RESPONSE_PAYLOAD (treatment arm, non-shadow) — versus being passed to the heuristic control. Denominator = DNN + heuristic (the two experiment arms); non-experiment traffic that can never reach the model is excluded.

Bars show daily transaction volume (right axis); lines show accuracy % (left axis).