gitadmin/foxhunt
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(aux): three bugs causing systematic anti-correlation (F1+F2)

Browse Source 
F-series investigations triangulated TWO real bugs that together
explained: aux BCE > ln(2) chance baseline, sub-chance dir_acc with
balanced labels, and bit-identical training across vastly different
POS_WEIGHT_MAX values.

BUG #1 (F2) — One-sided book contamination in mid_price_f32:
  - crates/ml-alpha/src/data/loader.rs::mid_price_f32 blindly averaged
    bid_px[0] + ask_px[0] without checking validity. MBP-10 snapshots
    at session boundaries / halts / stale level-0 vacancies have
    bid_px=0 or ask_px=0; result was mid = real_price/2 (~2750 for ES)
    or mid = 0 (both sides empty).
  - generate_outcome_labels_ab:218-220 only guarded is_finite() (NOT > 0).
    Sister fn generate_labels:75 does check both — asymmetry between
    parallel generators.
  - Transitions like mid=0 → mid=5500 produced ΔP=5500, instantly
    satisfying delta > 2×cost → spuriously triggering y_prof=1. Each
    contaminated snapshot tainted up to 2K downstream labels (appears
    as p_t for K positions and as p_kt for K positions).
  - With ~10-20% one-sided books in real ES, ~35-45% of K=10 labels
    spuriously positive — exactly matching the observed pos_fraction
    (0.35, 0.39, 0.46 for long).

  Fix (broader): mid_price_f32 returns NaN for one-sided/empty books
  at the source. Cascades to ALL downstream consumers (regime features,
  snap_features encoder, labels). Defensive guard also added in
  multi_horizon_labels.rs:218 for direct-test callers bypassing the
  loader.

BUG #2 (F1+F3) — dir_acc metric structural bias:
  - perception.rs:3647 match condition for flat-true bucket required
    float-EXACT equality on raw logits (`pred_diff == 0.0`) — essentially
    unreachable for continuous-valued logits.
  - On real ES, ~30-50% of samples are flat (neither direction
    profitable at 2×cost threshold). ALL counted as misses → random
    baseline depressed to ~0.35 (matching observed 0.32-0.45). BCE
    can DROP while dir_acc DROPS — decoupled metrics.
  - aux_lift_dir_acc_threshold=0.85 was structurally UNREACHABLE under
    this metric regardless of model quality.

  Fix: skip flat-true samples (`if true_diff == 0 { continue; }`).
  Converts dir_acc into "given a directional outcome, did we get the
  direction right?" with proper random baseline 0.5.

F3 + F4 confirmed: aux head wiring (8 head Adam optimizers, fwd/bwd
kernel arg order, BCE+sigmoid gradient sign, reduce_axis0 reduction)
is correct. The synthetic perception_overfit test stays in chance
regime because it constructs constant prof_long=1, prof_short=0 →
no zero-priced snapshots, no flat-true bucket. Production-data path
divergence is the canonical pattern in
pearl_canary_input_freshness_launch_order.

cargo check --workspace --all-targets: clean.
cargo test -p ml-alpha --lib multi_horizon_labels: 15 passed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-05-22 13:19:42 +02:00
parent f065cfbaa2
commit 85d3ca5c72
3 changed files with 38 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
crates/ml-alpha/src/data/loader.rs
View File

@@ -507,11 +507,26 @@ impl MultiHorizonLoader {
}
}
/// Mid price at level 0; returns NaN for one-sided / empty books per
/// the F2 finding (2026-05-22). MBP-10 snapshots at session boundaries,
/// halts, or stale level-0 vacancies have `bid_px = 0` or `ask_px = 0`;
/// the previous blind average produced `mid = real_price / 2 ≈ 2750`
/// (single side) or `mid = 0` (both sides empty). Downstream label
/// generation in `multi_horizon_labels::generate_outcome_labels_ab`
/// only guarded `is_finite()` (not `> 0`), so transitions from mid=0
/// to mid≈5500 produced `ΔP ≈ 5500`, spuriously satisfying the cost
/// threshold and inflating `pos_fraction` from the expected ~5% to
/// the observed ~40%. NaN return cascades through label generation
/// and the BCE NaN-mask handles it cleanly.
fn mid_price_f32(s: &Mbp10Snapshot) -> f32 {
let l = s.levels.first().copied().unwrap_or_else(BidAskPair::empty);
let bid = BidAskPair::price_to_f64(l.bid_px);
let ask = BidAskPair::price_to_f64(l.ask_px);
(0.5 * (bid + ask)) as f32
if bid > 0.0 && ask > 0.0 {
(0.5 * (bid + ask)) as f32
} else {
f32::NAN
}
}
fn convert(

8
crates/ml-alpha/src/multi_horizon_labels.rs
View File

@@ -215,7 +215,13 @@ pub fn generate_outcome_labels_ab(
for t in 0..n - k {
let p_t = prices[t];
let p_kt = prices[t + k];
if !p_t.is_finite() || !p_kt.is_finite() {
// F2 fix (2026-05-22): also reject p <= 0 to match `generate_labels:75`.
// mid_price_f32 in loader.rs now returns NaN for one-sided books, so
// the is_finite guard alone would suffice — this is belt+suspenders
// for callers that bypass the loader (e.g. direct unit tests passing
// raw price arrays). Without the `> 0` guard a transition from
// mid=0 to mid≈5500 produces ΔP=5500, spuriously triggering y_prof=1.
if !p_t.is_finite() || !p_kt.is_finite() || p_t <= 0.0 || p_kt <= 0.0 {
continue;
}
let delta = p_kt - p_t;

16
crates/ml-alpha/src/trainer/perception.rs
View File

@@ -3641,10 +3641,24 @@ impl PerceptionTrainer {
continue;
}
let true_diff = lt - st;
// F1 fix (2026-05-22): skip flat-true samples
// (`y_prof_long == y_prof_short`). On real ES data
// ~30-50% of samples are flat (neither direction
// profitable). The prior `true_diff == 0 &&
// pred_diff == 0.0` match condition required
// float-exact equality on raw logits — unreachable
// for continuous-valued logits, so all flat samples
// counted as misses and depressed dir_acc to ~0.35
// (matching observed 0.32-0.45). Skipping them
// converts dir_acc into "given a directional
// outcome, did we get the direction right?" with
// proper random baseline 0.5.
if true_diff == 0.0 {
continue;
}
let pred_diff = prof_long[idx] - prof_short[idx];
if (true_diff > 0.0 && pred_diff > 0.0)
|| (true_diff < 0.0 && pred_diff < 0.0)
|| (true_diff == 0.0 && pred_diff == 0.0)
{
per_h_match[h] += 1;
}