The Phase 1 multi-resolution layout (10 raw + 10 agg@30 + 12 agg@100) regressed ALL horizons in alpha-perception-htpp6 (2026-05-22): - auc_h100: 0.681 -> 0.512 (-0.169) - auc_h300: 0.617 -> 0.506 (-0.111) - auc_h1000: 0.576 -> 0.526 (-0.050) Hypothesis falsified. Root cause: Mamba2+CfC SSM encoder already used all 32 raw ticks effectively via state-recurrence; replacing 22 raw ticks with arithmetic-mean aggregates destroyed within-window microstructure variance (the actual h100 signal) AND broke temporal continuity that the recurrence relies on. Δt Fourier encoder couldn't compensate. Architectural pearl: SSM/RNN/CfC + multi-resolution input is incompatible without separate-encoder-per-scale or explicit scale tokens. Transformer-style positional encoding tolerates scale-mixing; recurrent state updates assume consecutive positions. Reverts default to '1:32'. Adds explicit single_scale_32() constructor for callers (harness, tests). Keeps default_three_scale() in code with deprecation note for future sub-variant experiments. Production defaults across alpha_train CLI, Argo template, dispatcher script, ml-backtesting harness now match the proven baseline.
114 lines
4.5 KiB
Rust
114 lines
4.5 KiB
Rust
//! Ring 1b — trainer/backtest parity check. See spec §8 Ring 1b.
|
|
//!
|
|
//! Verifies that the data-loading path used by the backtest harness
|
|
//! (`inference_only=true`) produces an Mbp10RawInput that is byte-equal
|
|
//! to what the trainer's loader produces from the same source.
|
|
//! Since both paths share the same loader code and same Mbp10Snapshot
|
|
//! → Mbp10RawInput conversion, the equality is structurally guaranteed —
|
|
//! but the test guards against any future refactor that breaks the claim.
|
|
//!
|
|
//! The full inference-output parity (asserting [probs; N_HORIZONS] from
|
|
//! the trunk's captured graph is bit-equal between training-time loader
|
|
//! and backtest-time loader paths) requires a real ml-alpha checkpoint,
|
|
//! which is deferred until a checkpoint-format is pinned in ml-alpha.
|
|
//! When a checkpoint is available, FOXHUNT_TEST_CKPT will gate the
|
|
//! gpu-inference assertion below.
|
|
|
|
use anyhow::Result;
|
|
use ml_alpha::data::loader::{
|
|
discover_mbp10_files_sorted, InstrumentFilter, MultiHorizonLoader, MultiHorizonLoaderConfig,
|
|
};
|
|
use std::path::PathBuf;
|
|
|
|
fn try_loader(inference_only: bool) -> Option<MultiHorizonLoader> {
|
|
let root = std::env::var("FOXHUNT_TEST_DATA").ok()?;
|
|
let mbp10 = PathBuf::from(&root).join("ES.FUT");
|
|
if !mbp10.exists() {
|
|
return None;
|
|
}
|
|
let files = discover_mbp10_files_sorted(&mbp10).ok()?;
|
|
let cfg = MultiHorizonLoaderConfig {
|
|
files,
|
|
predecoded_dir: mbp10.clone(),
|
|
multi_resolution: ml_alpha::data::aggregation::MultiResolutionConfig::single_scale_32(),
|
|
horizons: ml_alpha::heads::HORIZONS,
|
|
n_max_sequences: 1,
|
|
seed: 0xCAFEF00D,
|
|
inference_only,
|
|
outcome_label_cost: ml_alpha::data::loader::DEFAULT_OUTCOME_LABEL_COST_ES,
|
|
instrument_filter: InstrumentFilter::All,
|
|
};
|
|
match MultiHorizonLoader::new(&cfg) {
|
|
Ok(l) => Some(l),
|
|
Err(e) => {
|
|
eprintln!("skipping: fixture data not usable ({e})");
|
|
None
|
|
}
|
|
}
|
|
}
|
|
|
|
#[test]
|
|
#[ignore = "requires populated FOXHUNT_TEST_DATA"]
|
|
fn peek_first_byte_equal_across_modes() -> Result<()> {
|
|
let Some(loader_train) = try_loader(false) else { return Ok(()); };
|
|
let Some(loader_bt) = try_loader(true) else { return Ok(()); };
|
|
|
|
let first_train = loader_train.peek_first()?;
|
|
let first_bt = loader_bt.peek_first()?;
|
|
|
|
// Inputs must be byte-equal (Mbp10RawInput is Pod via #[derive(Default)]
|
|
// — bid_px/sz/ask_px/sz arrays of f32, plus a handful of scalars).
|
|
assert_eq!(first_train.ts_ns, first_bt.ts_ns);
|
|
assert_eq!(first_train.prev_ts_ns, first_bt.prev_ts_ns);
|
|
for k in 0..10 {
|
|
assert_eq!(
|
|
first_train.bid_px[k].to_bits(),
|
|
first_bt.bid_px[k].to_bits(),
|
|
"bid_px[{k}] bit-mismatch between training and inference loader"
|
|
);
|
|
assert_eq!(first_train.bid_sz[k].to_bits(), first_bt.bid_sz[k].to_bits());
|
|
assert_eq!(first_train.ask_px[k].to_bits(), first_bt.ask_px[k].to_bits());
|
|
assert_eq!(first_train.ask_sz[k].to_bits(), first_bt.ask_sz[k].to_bits());
|
|
}
|
|
for k in 0..6 {
|
|
assert_eq!(
|
|
first_train.regime[k].to_bits(),
|
|
first_bt.regime[k].to_bits(),
|
|
"regime[{k}] bit-mismatch"
|
|
);
|
|
}
|
|
assert_eq!(first_train.prev_mid.to_bits(), first_bt.prev_mid.to_bits());
|
|
assert_eq!(first_train.trade_signed_vol.to_bits(), first_bt.trade_signed_vol.to_bits());
|
|
assert_eq!(first_train.trade_count, first_bt.trade_count);
|
|
|
|
Ok(())
|
|
}
|
|
|
|
#[test]
|
|
#[ignore = "requires populated FOXHUNT_TEST_DATA"]
|
|
fn inference_iteration_matches_chronological_snapshots() -> Result<()> {
|
|
let Some(mut loader) = try_loader(true) else { return Ok(()); };
|
|
|
|
// Pull 5 inference inputs; verify timestamps are monotone non-decreasing
|
|
// and each input's prev_ts_ns equals the previous input's ts_ns
|
|
// (after the first, which uses cur=prev semantics).
|
|
let mut prev_ts_ns = 0u64;
|
|
let mut first_iter = true;
|
|
for i in 0..5 {
|
|
let snap = loader.next_inference_input()?.expect("stream not exhausted");
|
|
if first_iter {
|
|
// peek_first/next-first uses cur==prev → prev_ts == ts.
|
|
assert_eq!(snap.prev_ts_ns, snap.ts_ns, "first snap has prev_ts == ts");
|
|
first_iter = false;
|
|
} else {
|
|
assert_eq!(
|
|
snap.prev_ts_ns, prev_ts_ns,
|
|
"snap {i}: prev_ts_ns should equal previous snap's ts_ns"
|
|
);
|
|
}
|
|
assert!(snap.ts_ns >= prev_ts_ns, "snap {i}: monotonic ts violated");
|
|
prev_ts_ns = snap.ts_ns;
|
|
}
|
|
Ok(())
|
|
}
|