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

fix(loader): rename EMA bandwidth constants to reflect actual half-lives

Browse Source 
Empirical ES MBP-10 median inter-event Δt is 74µs (not 250ms as
documented in loader.rs:25-26). The constants ALPHA_MED=0.02 and
ALPHA_SLOW=0.0005 were named/commented as "9s @ 250ms" and "6min @
250ms" but at the actual 74µs Δt their half-lives are ~2.6ms and ~100ms
— a 3000× scale mismatch.

Rename for truth-in-naming (values unchanged to preserve all trained
checkpoints):
- ALPHA_MED → ALPHA_FAST_2P6MS (half-life ~2.6ms @ 74µs Δt)
- ALPHA_SLOW → ALPHA_MED_100MS (half-life ~100ms @ 74µs Δt)

The regime[0..5] feature naming in snap_features.rs reflects the
actual microstructure-burst and 1-2-second-clustering time scales, NOT
macro 9s/6min regimes. True macro-regime EMAs (if needed) are a
separate addition.

ml-alpha lib: 33 passed, unchanged from baseline.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
jgrusewski
2026-05-22 00:08:20 +02:00
parent e18325aaf1
commit fe5d7a1ad7
2 changed files with 49 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
crates/ml-alpha/src/cfc/snap_features.rs
View File

@@ -22,16 +22,19 @@ pub const FEATURE_DIM: usize = 40;
/// Number of regime-context features pre-computed by the loader and
/// piped into snap_features slots `out[20..20+REGIME_DIM]`. These give
/// the model multi-minute trend / volatility / liquidity context that
/// is structurally unreachable inside the K-snapshot BPTT window.
/// the model microstructure-burst (~2.6ms) and 1-2-second-clustering
/// (~100ms) trend / volatility / liquidity context that is structurally
/// unreachable inside the K-snapshot BPTT window. (NOTE: these are NOT
/// macro-minute regimes — see `loader.rs` ALPHA_* constants for the
/// empirically-derived half-lives at the actual 74µs median Δt.)
///
/// Slot assignments (loader computes via EMA cascade):
/// regime[0] = (mid - ema_mid_med) / sqrt(ema_rv_med + eps) z-score, ~35s
/// regime[1] = (mid - ema_mid_slow) / sqrt(ema_rv_slow + eps) z-score, ~6min
/// regime[2] = (ema_mid_med - ema_mid_slow) / sqrt(ema_rv_slow + eps) trend signal
/// regime[3] = log1p(ema_rv_slow * 1e4) log-compressed slow vol
/// regime[4] = log1p(ema_spread_in_ticks_med) liquidity regime
/// regime[5] = log1p(ema_trades_per_snap_med * 100) activity regime
/// regime[0] = (mid - ema_mid_fast) / sqrt(ema_rv_fast + eps) z-score, ~2.6ms half-life
/// regime[1] = (mid - ema_mid_med) / sqrt(ema_rv_med + eps) z-score, ~100ms half-life
/// regime[2] = (ema_mid_fast - ema_mid_med) / sqrt(ema_rv_med + eps) trend signal
/// regime[3] = log1p(ema_rv_med * 1e4) log-compressed med vol
/// regime[4] = log1p(ema_spread_in_ticks_fast) liquidity regime
/// regime[5] = log1p(ema_trades_per_snap_fast * 100) activity regime
pub const REGIME_DIM: usize = 6;
/// One raw MBP-10 snapshot in struct-of-arrays form. Phase A loader

60
crates/ml-alpha/src/data/loader.rs
View File

@@ -22,8 +22,24 @@ use crate::cfc::snap_features::{Mbp10RawInput, ES_TICK_SIZE, REGIME_DIM};
use crate::multi_horizon_labels::generate_labels;
/// EMA smoothing factors. Half-life ≈ ln(2)/α snapshots.
const ALPHA_MED: f32 = 0.02; // half-life ≈ 35 snapshots ≈ 9s @ 250ms
const ALPHA_SLOW: f32 = 0.0005; // half-life ≈ 1400 snapshots ≈ 6 min
///
/// Empirical ES MBP-10 median inter-event Δt = 74µs (NOT 250ms as the
/// previous comments claimed — mean Δt is 55ms but the distribution is
/// extremely heavy-tailed so the median dominates EMA dynamics). At that
/// rate the half-lives in wall-clock are roughly:
///
/// * `ALPHA_FAST_2P6MS` (α=0.02) → ~35 snapshots × 74µs ≈ 2.6ms
/// — microstructure-burst scale
/// * `ALPHA_MED_100MS` (α=0.0005) → ~1400 snapshots × 74µs ≈ 100ms
/// — 1-2-second clustering scale
///
/// Values preserved (validated by all prior training); the names are
/// updated to reflect what they actually do. True macro-regime EMAs
/// (multi-second / multi-minute half-lives) would need α ~5e-6 or
/// smaller and are deliberately NOT added here — that's a separate
/// feature-engineering decision.
const ALPHA_FAST_2P6MS: f32 = 0.02;
const ALPHA_MED_100MS: f32 = 0.0005;
const REGIME_EPS: f32 = 1e-6;
/// Compute per-snapshot regime features via a single forward EMA pass
@@ -39,10 +55,10 @@ fn compute_regime_features(snapshots: &[Mbp10Snapshot]) -> Vec<[f32; REGIME_DIM]
let mut out: Vec<[f32; REGIME_DIM]> = vec![[0.0; REGIME_DIM]; n];
if n == 0 { return out; }
let mut ema_mid_fast = f32::NAN;
let mut ema_mid_med = f32::NAN;
let mut ema_mid_slow = f32::NAN;
let mut ema_mid2_med = f32::NAN; // E[mid²] for var estimate
let mut ema_mid2_slow = f32::NAN;
let mut ema_mid2_fast = f32::NAN; // E[mid²] for var estimate
let mut ema_mid2_med = f32::NAN;
let mut ema_spread = f32::NAN; // ticks
let mut ema_trades = f32::NAN; // per-snapshot delta count
let mut prev_trade_count: u32 = 0;
@@ -57,30 +73,30 @@ fn compute_regime_features(snapshots: &[Mbp10Snapshot]) -> Vec<[f32; REGIME_DIM]
prev_trade_count = snap.trade_count;
// Update EMAs with sentinel-bootstrap.
if ema_mid_med.is_nan() {
ema_mid_med = mid; ema_mid_slow = mid;
ema_mid2_med = mid * mid; ema_mid2_slow = mid * mid;
if ema_mid_fast.is_nan() {
ema_mid_fast = mid; ema_mid_med = mid;
ema_mid2_fast = mid * mid; ema_mid2_med = mid * mid;
ema_spread = spread_t; ema_trades = trades as f32;
} else {
ema_mid_med = (1.0 - ALPHA_MED) * ema_mid_med + ALPHA_MED * mid;
ema_mid_slow = (1.0 - ALPHA_SLOW) * ema_mid_slow + ALPHA_SLOW * mid;
ema_mid2_med = (1.0 - ALPHA_MED) * ema_mid2_med + ALPHA_MED * mid * mid;
ema_mid2_slow = (1.0 - ALPHA_SLOW) * ema_mid2_slow + ALPHA_SLOW * mid * mid;
ema_spread = (1.0 - ALPHA_MED) * ema_spread + ALPHA_MED * spread_t;
ema_trades = (1.0 - ALPHA_MED) * ema_trades + ALPHA_MED * trades as f32;
ema_mid_fast = (1.0 - ALPHA_FAST_2P6MS) * ema_mid_fast + ALPHA_FAST_2P6MS * mid;
ema_mid_med = (1.0 - ALPHA_MED_100MS) * ema_mid_med + ALPHA_MED_100MS * mid;
ema_mid2_fast = (1.0 - ALPHA_FAST_2P6MS) * ema_mid2_fast + ALPHA_FAST_2P6MS * mid * mid;
ema_mid2_med = (1.0 - ALPHA_MED_100MS) * ema_mid2_med + ALPHA_MED_100MS * mid * mid;
ema_spread = (1.0 - ALPHA_FAST_2P6MS) * ema_spread + ALPHA_FAST_2P6MS * spread_t;
ema_trades = (1.0 - ALPHA_FAST_2P6MS) * ema_trades + ALPHA_FAST_2P6MS * trades as f32;
}
let var_fast = (ema_mid2_fast - ema_mid_fast * ema_mid_fast).max(0.0);
let var_med = (ema_mid2_med - ema_mid_med * ema_mid_med ).max(0.0);
let var_slow = (ema_mid2_slow - ema_mid_slow * ema_mid_slow).max(0.0);
let std_fast = (var_fast + REGIME_EPS).sqrt();
let std_med = (var_med + REGIME_EPS).sqrt();
let std_slow = (var_slow + REGIME_EPS).sqrt();
out[k][0] = (mid - ema_mid_med) / std_med; // z-score med
out[k][1] = (mid - ema_mid_slow) / std_slow; // z-score slow
out[k][2] = (ema_mid_med - ema_mid_slow) / std_slow; // trend signal
out[k][3] = (1.0 + var_slow.sqrt() * 1e4).ln(); // log-compressed slow vol
out[k][4] = (1.0 + ema_spread).ln(); // liquidity regime
out[k][5] = (1.0 + ema_trades * 100.0).ln(); // activity regime
out[k][0] = (mid - ema_mid_fast) / std_fast; // z-score fast (~2.6ms)
out[k][1] = (mid - ema_mid_med) / std_med; // z-score med (~100ms)
out[k][2] = (ema_mid_fast - ema_mid_med) / std_med; // trend signal
out[k][3] = (1.0 + var_med.sqrt() * 1e4).ln(); // log-compressed med vol
out[k][4] = (1.0 + ema_spread).ln(); // liquidity regime
out[k][5] = (1.0 + ema_trades * 100.0).ln(); // activity regime
}
out
}