ColdFire / Model Research

Model Candidate Search

Candidate queue for testing model families on top of Feature Experiment Model Experiment Version 1. The goal is to compare AUC, OOS policy PnL, and 72h/96h roleplay deltas before touching production.

Baseline

Experiment Rules

A model is deployable only if it beats current v1 on more than one axis. A pure PnL win without AUC/calibration support is treated as p-hacking risk.

Training Sample Logic

Decision-Time Match

P5 tests change only row selection or row weights. Features, LightGBM family, and 120s policy stay fixed.

Results So Far

Use the sort controls to inspect raw AUC, dAUC, OOS PnL, and dPnL. A candidate is not considered better unless both AUC and OOS policy PnL improve; roleplay is a stability guardrail.
Rank Model Status AUC dAUC OOS PnL dPnL 72h Roleplay 96h Roleplay Read

Sequential Sample Tests

Model Status Train Rows Flips Full dAUC Full dPnL Seq dPnL Fixed Attempt dAUC 96h Roleplay

Calibration-Only Tests

Model Status Cal Rows Flips Full dAUC Full dPnL Seq dPnL Fixed Attempt dAUC 96h Roleplay

P8 Iteration Winners

Candidate Logic AUC Delta PnL Delta Seq PnL Delta 72h Roleplay 96h Roleplay Orders 96h Losses Rejected 96h Wins Rejected Read

P9 Holdout Check

Candidate dAUC dPnL Early OOS dPnL Late OOS dPnL Older Roleplay Recent 96h Recent 72h Robust? Read

P10/P11 Reweighting

Candidate Rule dAUC dPnL Seq dPnL Fixed Attempt dAUC 72h Roleplay 96h Roleplay 168h Roleplay Read

Run Queue

Priority Status Model Variant Dependency Why test it Risk dAUC dPnL dRoleplay

Phases

Research Sources