jgrusewski d42a8f4000 fix: reward-scale v_range floor + fold-boundary shrink-and-perturb ...

Breaks C51 stable fixed point: tight v_range → tiny Bellman shift → tiny gradient
→ Q-values stuck near zero → v_range stays tight (self-reinforcing trap).

- REWARD_SCALE_FLOOR=0.01 replaces ABS_FLOOR=1e-4 (was 1400× too small)
- v_range floor proportional to single-step reward magnitude, not Q-statistics
- Ensures Bellman projection can shift atoms by at least 1 reward unit
- Removes pessimistic Q-init bias (incompatible with adaptive v_range)

Fold transition stability:
- Shrink-and-perturb at fold boundary (alpha=0.8, sigma=0.01)
- Reduces overfit to previous fold's data distribution
- Eliminates val_Sharpe=-33 crash on fold 2/3 transitions

Result: val_Sharpe positive across all 50 epochs and 3 walk-forward folds.
Q-values grow monotonically (0→0.023), Q-gap grows 245× (2.7e-6→6.6e-4).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

2026-04-12 20:25:38 +02:00

..

backtesting

refactor: collapse 9-level to 7-level ExposureLevel — eliminate degenerate Flat variants

2026-04-11 11:54:09 +02:00

common

fix: switch tracing to stderr (unbuffered) — stdbuf has no effect on Rust

2026-04-02 16:29:55 +02:00

config

…

ctrader-openapi

fix: tune log levels across workspace — demote noisy warn to debug/trace

2026-03-14 11:35:15 +01:00

data

…

database

…

market-data

…

ml

fix: reward-scale v_range floor + fold-boundary shrink-and-perturb

2026-04-12 20:25:38 +02:00

ml-asset-selection

…

ml-backtesting

…

ml-checkpoint

…

ml-core

cleanup: remove ALL GpuVarStore from DQN path + re-enable bottleneck_dim=16

2026-04-11 14:42:54 +02:00

ml-data

…

ml-data-validation

…

ml-dqn

WIP: training quality fixes + VarStore elimination (INCOMPLETE — needs xavier_init_params_buf)

2026-04-11 21:03:05 +02:00

ml-ensemble

feat(ensemble): GPU-native ensemble reduce kernel, remove CPU fallbacks

2026-03-29 22:35:37 +02:00

ml-explainability

feat(ensemble): GPU-native ensemble reduce kernel, remove CPU fallbacks

2026-03-29 22:35:37 +02:00

ml-features

test(data): per-file vs global front-month filter — documents quarterly roll bug

2026-04-08 21:31:46 +02:00

ml-hyperopt

refactor: LHS is the correct algorithm for small budgets, not a fallback

2026-03-29 19:18:47 +02:00

ml-labeling

fix(clippy): ml-supervised + ml-core + thin crates — 34 errors fixed

2026-03-19 00:39:03 +01:00

ml-observability

…

ml-paper-trading

…

ml-ppo

refactor: collapse 9-level to 7-level ExposureLevel — eliminate degenerate Flat variants

2026-04-11 11:54:09 +02:00

ml-regime

refactor(cuda): eliminate all CPU fallbacks — CUDA mandatory across ML stack

2026-03-16 21:01:28 +01:00

ml-regime-detection

…

ml-risk

…

ml-security

…

ml-stress-testing

…

ml-supervised

cleanup: remove half crate dependency from entire workspace

2026-04-10 18:54:14 +02:00

ml-universe

…

ml-validation

…

model_loader

…

risk

refactor: eliminate candle from entire workspace — tests, examples, Cargo.toml

2026-03-18 00:53:47 +01:00

risk-data

fix: tune log levels across workspace — demote noisy warn to debug/trace

2026-03-14 11:35:15 +01:00

storage

…

trading_engine

fix(ml): BF16/F32 boundary alignment, GPU-resident ops across all ML crates

2026-03-15 11:59:31 +01:00

trading-data

…

training_uploader

…