Loads trained DQN/PPO checkpoints, runs greedy inference on walk-forward
test windows, computes Sharpe ratio, max drawdown, win rate, profit factor,
and total return per fold. Outputs a JSON evaluation report with aggregate
metrics and sanity checks (beats-random, action diversity, fold consistency).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add ml/examples/train_baseline.rs that trains DQN and PPO models using
expanding walk-forward windows on real Databento OHLCV data.
Features:
- CLI args via clap (--model, --epochs, --batch-size, --data-dir, etc.)
- Recursive .dbn.zst file discovery and OHLCV bar loading
- 51-dim feature extraction via extract_ml_features()
- Walk-forward window generation with NormStats per fold
- DQN training loop with epsilon-greedy, experience replay, early stopping
- PPO training loop with GAE, trajectory collection, early stopping
- PnL-based reward (BUY/SELL/HOLD)
- Safetensors checkpoint saving per fold
- NormStats JSON export for evaluation reproducibility
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Implements expanding-window walk-forward cross-validation for time-series
ML models, preventing lookahead bias by always evaluating on unseen future
data. Includes z-score NormStats computed from training splits only.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds `ml/examples/download_baseline.rs` that downloads 730 days of
Databento OHLCV-1m data in quarterly chunks for 4 CME futures symbols.
Features:
- Reads universe config from TOML (symbols, date range, dataset)
- Splits date range into calendar-quarter chunks (~90 days each)
- Resume support: skips existing non-empty files
- Uses `get_range_to_file` for streaming writes to .dbn.zst
- Dry-run mode with cost estimate ($0.12/symbol/day)
- Confirmation prompt (skippable with --yes)
- Per-file progress with timing and byte counts
- Failure-tolerant: logs errors and continues
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Sequence-buffered adapter with 3-layer MLP projection
(flat_dim -> hidden -> hidden/2 -> 1). Ring buffer collects
seq_len feature vectors, returns neutral prediction until full,
then flattens and projects through the MLP with sigmoid output
mapping to direction [-1,1] and confidence [0,1].
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Remove panic!() calls from test_futures_baseline_micro_mapping,
use map()+Some() pattern consistent with rest of test suite
- Make DatasetSpec::from_universe() accept a name parameter instead
of hardcoding "futures-baseline"
- Add doc comment to UniverseConfigMeta explaining dead_code fields
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Remove duplicate AssetClass enum from data_pipeline module and replace
with a re-export from asset_selection, which is now the canonical
location. This ensures data_pipeline::AssetClass and
asset_selection::AssetClass are the same type, enabling direct
comparison and preventing subtle type mismatch bugs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Gate optimizer adjusts conviction gate thresholds based on win-rate per
confidence bucket with cooldown and kill switch safety rails. Model
registry provides lifecycle management (Candidate → Staging → Production
→ Archived) with InMemoryModelRegistry for testing. P&L attribution
decomposes realized trade P&L into per-model contributions using signal
alignment.
24 new tests across 3 modules.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Introduces the asset_selection module with AssetClass enum (Equity/ETF/Future),
UniverseAsset struct, and AssetUniverse with filtering, lookup, and a us_starter()
preset of 7 highly liquid US assets. Includes scorer/selector placeholder stubs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adjusts model weights based on rolling Sharpe ratios with:
- EMA smoothing (alpha=0.1)
- Bounds: [0.05, 0.40] per model
- Max step: 0.03 per cycle
- 24h cooldown, 7-day grace period for new models
- Kill switch freeze/unfreeze
8 tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds optional ConvictionGateEvaluator field to EnsembleCoordinator.
When configured, predict() evaluates all 7 gates before returning.
If any gate rejects, returns HOLD with rejection metadata.
Includes quorum ratio, trading session, and regime volatility helpers.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
7-gate conviction system, autonomous feedback loop with kill switch,
and Rust-native model registry — backed by PostgreSQL + QuestDB.
Resolve merge conflicts in hyperopt/adapters/mod.rs and enhanced_ml.rs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Scale model_memory_mb by activation_multiplier in resolve_batch_size()
so TFT (2.5x) gets proportionally smaller batches than DQN (1.0x)
- Add cached_capabilities() with OnceLock to avoid spawning nvidia-smi
on every call to CampaignConfig::dqn_default/ppo_default/PpoTrainer::new
- Add PartialEq to GpuCapabilities and simplify serde roundtrip test
- Update stale "RTX 3050 Ti" and "Exceeds GPU limit (230)" comments
to reflect dynamic detection
- Document Auto variant silent CPU fallback behavior (no callers affected)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add DeviceConfig and GpuCapabilities re-exports to ml::prelude so
downstream crates can import GPU management types via a single
`use ml::prelude::*` statement.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace hardcoded max_batch_size: 230 in CampaignConfig::dqn_default()
and ppo_default() with dynamic GPU detection via GpuCapabilities::detect()
and resolve_batch_size(). Update test to no longer assert <= 230.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace Device::cuda_if_available(0).unwrap_or(Device::Cpu) with
DeviceConfig::Auto.resolve().unwrap_or(Device::Cpu) in DQN, PPO, TFT,
and Mamba2 ensemble inference adapters so device selection goes through
the centralized GPU detection module.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add check_gradients_finite() to gradient_utils.rs that detects NaN/Inf
in GradStore before optimizer step. Uses efficient sum_all approach
where any NaN element produces a NaN sum. Includes 3 unit tests
(finite pass, NaN detected, empty vars pass).
DQN and PPO training loops already wired via gradient_accumulation
module's check_gradients_finite (all 6 paths: standard policy/value,
remainder policy/value, LSTM policy/value).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add check_gradients_finite() utility that scans GradStore for NaN/Inf
values. Wire into DQN (after gradient accumulation, before optimizer step)
and PPO (both MLP and LSTM training paths, after backward pass).
Prevents silent model corruption from exploding gradients or numerical
instability — training halts immediately with a descriptive error.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace local DeviceConfig enum definition in liquid/candle_cfc.rs with
a re-export from the central crate::gpu::DeviceConfig, eliminating
duplication while preserving all existing API and tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add optional correlation matrix parameter to mean-variance optimization.
When provided, builds full covariance matrix (Sigma[i][j] = corr[i][j] *
vol_i * vol_j) instead of diagonal-only. Existing API unchanged — callers
pass None by default. New allocate_with_correlations() public method for
correlated optimization. Five new tests: identity-matches-diagonal,
correlated-differs-from-diagonal, invalid dimensions, non-square matrix,
and non-MeanVariance delegation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- NoiseScheduler: precomputed cosine/linear alpha_bar schedules
- Denoiser: FC network with sinusoidal time embedding + SiLU + residual
- DDIMSampler: deterministic fast sampling (10 steps from 1000 timesteps)
- DiffusionTrainableAdapter: UnifiedTrainable for unified training pipeline
- Hyperopt adapter with ParameterSpace (9 params, batch ≤64 for 4GB GPU)
- ModelType::Diffusion registered in common + coordinator
- 41 tests passing (config=3, noise=7, denoiser=4, sampler=5, trainable=12, hyperopt=7)
- OOM-safe: FC denoiser instead of U-Net, small hidden dims, conservative defaults
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Phase 2-3 of ensemble expansion:
- KAN (Kolmogorov-Arnold Network): B-spline basis, layer, network, trainable adapter
- TLOB UnifiedTrainable adapter with 3D input support (batch, seq, features)
- Hyperopt adapters for both KAN and TLOB (ParameterSpace + metrics)
- ModelType::KAN variant registered in common, coordinator, lib.rs
- 44 new tests, all passing, zero warnings
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Closes the integration gap between Liquid CfC and DQN/PPO by adding:
- LiquidTrainer with gRPC progress callbacks, early stopping, and
checkpoint management (ml/src/trainers/liquid.rs)
- LiquidModel wrapper in enhanced_ml.rs for hot-loading from safetensors
- Ensemble weight rebalance: DQN 0.25, PPO 0.25, TFT 0.20, Mamba2 0.15,
Liquid-CfC 0.15
Verified: 81 liquid unit tests + 3 integration tests + 211 trading_service
tests pass, 0 compile errors across workspace.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
