Эксперимент: Measure-Valued Forecast Martingale Regularizer (#1250)

{ "worked": false, "confidence": 8, "verdict": "Built a PyTorch 6-step GRU categorical forecaster with the finite-feature measure-valued martingale regularizer, plus a direct toy verification. The toy check reproduced the claimed phenomenon: a history-biased increment had a 436x larger moment norm than an unbiased increment. However, lambda values in the proposed range 0.01–1 produced essentially no change in held-out violations, while lambda=1000 gave only a modest and inconsistent reduction, without reliable calibration improvement.", "metrics": { "baseline": "lambda=0: final NLL 0.2993, accuracy 89.45%, final ECE 0.0487, mean held-out moment norm 0.01976", "idea": "lambda=0.1: final NLL 0.2993, accuracy 89.45%, final ECE 0.0487, mean held-out moment norm 0.01976; lambda=1: NLL 0.2993, accuracy 89.45%, ECE 0.0487, norm 0.01974; lambda=1000 primary seed: NLL 0.2875, accuracy 89.71%, ECE 0.0499, norm 0.01688" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 martingale_experiment.py", "files": [ "martingale_experiment.py", "results.json" ], "limitations": "The dataset is synthetic rather than revealed MNIST or CIFAR, training uses full-batch optimization, and only a small number of seeds was tested. Corruption or shift robustness, bootstrap confidence intervals, FLOP accounting, and larger recurrent or state-space models were not evaluated." }