feat: measurement's first deflatable metric — the Information Coefficient

Give a measurement run a standalone post-run quality score: the Information
Coefficient (IC), corr(signal_t, forward_return_{t+h}), with a permutation
null model. Before this a measurement run persisted only tap names and series
(MeasurementReport) — inspectable but not rankable or deflatable. IC is the
first metric measurement supplies: the tap-to-scalar reduction the C28
process-column / metric-interface seam anticipates ("a named metric vocabulary
supplied by measurement and backtest instead of baked in R-only").

Surface: `aura measure ic <run> --signal <tap> --price <tap> [--horizon]
[--permutations] [--seed]`. It reads the run's two recorded tap traces, builds
the forward-return series offline over the recorded price spine, aligns
signal -> return by exact timestamp, computes the IC, and prints it with a
one-sided permutation-null overfit probability (Laplace-smoothed, mirroring the
R deflation arm's formula in optimize_deflated).

Placement follows the C28 ladder: the generic pieces — pearson_corr and a
Fisher-Yates permute over SplitMix64 — land in aura-analysis (the domain-free
statistics foundation, beside resample_block); the IC semantics (forward-return
horizon, the offline join) and the CLI verb land in aura-cli (the shell). The
run path (run_measurement), MeasurementReport, the trace store, and the registry
metric vocabulary are untouched — existing `aura run` output stays byte-identical
(the C18 golden run_prints_json_and_exits_zero stays green).

Causality (C2): the forward-return read close_{t+h} is over a completed run's
recorded trace, not an in-graph node. A look-ahead signal is structurally
impossible in a causal run, so the engineered-vs-noise acceptance property is
unit-tested over hand-built offline series (perfect signal -> IC ~= 1, p < 0.05;
a varying but exactly-uncorrelated signal -> IC = 0, not significant;
constant/degenerate -> the (0.0, 1.0) floor), while the E2E validates the CLI
wiring, report well-formedness, determinism, and the error paths over a real run.

This lands the second, structurally distinct null-model computation (permutation,
beside the moving-block bootstrap behind R). That makes the deferred registry
deflation-vocabulary abstraction (#147 item 2) demand-driven rather than
speculative — the next cycle.

Verified: cargo test --workspace green (incl. the new aura-analysis unit tests,
the aura-cli ic_* unit tests, and the measure_ic E2E); cargo clippy --workspace
--all-targets -D warnings clean; the C18 byte-identity golden unchanged.

closes #290
refs #147
This commit is contained in:
2026-07-20 16:27:57 +02:00
parent a56ab7859d
commit a9d36ddd70
5 changed files with 458 additions and 0 deletions
+265
View File
@@ -35,6 +35,7 @@ use aura_backtest::{
RunMetrics, RunReport, SimBroker, SweepFamily, SweepPoint, WalkForwardResult, WindowRun,
PM_FIELD_NAMES, PM_RECORD_KINDS,
};
use aura_analysis::{pearson_corr, permute, SplitMix64};
use aura_std::{
GatedRecorder, LinComb, Recorder, RollingMax, RollingMin, SeriesReducer, Sub,
};
@@ -1843,6 +1844,88 @@ fn measurement_manifest(
}
}
/// The pure IC reduction result (no CLI context) — unit-testable in isolation.
/// Exercised by the in-crate `ic_tests` module and consumed by the run/command-path
/// wiring in `dispatch_measure_ic`, which constructs `IcReport` from a live
/// measurement run's recorded taps.
struct IcOutcome {
information_coefficient: f64,
overfit_probability: f64,
n_pairs: usize,
}
/// One measurement run's IC scalar + its permutation-null significance, for stdout.
#[derive(serde::Serialize)]
struct IcReport {
run: String,
signal_tap: String,
price_tap: String,
horizon: usize,
permutations: usize,
seed: u64,
n_pairs: usize,
information_coefficient: f64,
overfit_probability: f64,
}
impl IcReport {
fn to_json(&self) -> String {
serde_json::to_string(self).expect("a finite IcReport always serializes")
}
}
/// `corr(signal_t, forward_return_{t+h})` over two recorded tap series, with a
/// permutation null. `forward_return[i] = (price[i+h] - price[i]) / price[i]` on the
/// price ts-spine (a post-run array shift over recorded data — C2 governs in-graph
/// nodes, not a completed run's trace); signal is aligned to it by EXACT timestamp
/// match (unmatched dropped). `< 2` aligned pairs or zero variance → the degenerate
/// floor `(0.0, 1.0)`. One-sided null (higher-is-better), Laplace-smoothed like the
/// R deflation arm.
fn information_coefficient(
signal: &[(Timestamp, f64)],
price: &[(Timestamp, f64)],
horizon: usize,
permutations: usize,
seed: u64,
) -> IcOutcome {
use std::collections::HashMap;
// signal value by timestamp-i64 (`Timestamp` is a tuple struct over epoch-ns; a tap
// fires at most once per ts). `t.0` is the epoch-ns i64.
let sig_at: HashMap<i64, f64> = signal.iter().map(|(t, v)| (t.0, *v)).collect();
// forward returns on the price spine, paired with the signal at the SAME ts
let (mut sigs, mut frs) = (Vec::new(), Vec::new());
if horizon >= 1 && price.len() > horizon {
for i in 0..price.len() - horizon {
let (t, p0) = (price[i].0, price[i].1);
if p0 == 0.0 {
continue; // undefined return
}
let fr = (price[i + horizon].1 - p0) / p0;
if let Some(&s) = sig_at.get(&t.0) {
sigs.push(s);
frs.push(fr);
}
}
}
let n_pairs = sigs.len();
if n_pairs < 2 {
return IcOutcome { information_coefficient: 0.0, overfit_probability: 1.0, n_pairs };
}
let raw = pearson_corr(&sigs, &frs);
// permutation null: shuffle signal against fixed returns, one-sided count
let mut rng = SplitMix64::new(seed);
let mut perm = sigs.clone();
let mut ge = 0usize;
for _ in 0..permutations {
permute(&mut perm, &mut rng);
if pearson_corr(&perm, &frs) >= raw {
ge += 1;
}
}
let overfit_probability = (ge + 1) as f64 / (permutations as f64 + 1.0);
IcOutcome { information_coefficient: raw, overfit_probability, n_pairs }
}
/// The bare measurement run (C28 phase 3): `run_signal_r` MINUS `wrap_r` and the
/// eq/ex/r R-evaluation, KEEPING the declared-tap bind → drain → persist (C27).
/// No broker, no risk executor, no per-cycle equity/exposure/r recorders — this
@@ -2968,6 +3051,8 @@ enum Command {
Campaign(research_docs::CampaignCmd),
/// Inspect the project's data archive.
Data(DataCmd),
/// Reduce a measurement run's recorded taps to a signal-quality metric.
Measure(MeasureCmd),
}
#[derive(Args)]
@@ -3032,6 +3117,34 @@ struct DataCoverageCmd {
symbol: String,
}
#[derive(Args)]
struct MeasureCmd {
#[command(subcommand)]
command: MeasureCommand,
}
#[derive(Subcommand)]
enum MeasureCommand {
/// Information Coefficient of a signal tap against forward returns of a price tap.
Ic(MeasureIcCmd),
}
#[derive(Args)]
struct MeasureIcCmd {
/// The persisted run name (its trace-store subdirectory).
run: String,
#[arg(long)]
signal: String,
#[arg(long)]
price: String,
#[arg(long, default_value_t = 1)]
horizon: usize,
#[arg(long, default_value_t = 1000)]
permutations: usize,
#[arg(long, default_value_t = 0)]
seed: u64,
}
#[derive(Args)]
#[command(args_conflicts_with_subcommands = true)]
struct GraphCmd {
@@ -4012,6 +4125,46 @@ fn dispatch_data(cmd: DataCmd, env: &project::Env) {
}
}
fn dispatch_measure(cmd: MeasureCmd, env: &project::Env) {
match cmd.command {
MeasureCommand::Ic(a) => dispatch_measure_ic(a, env),
}
}
fn dispatch_measure_ic(a: MeasureIcCmd, env: &project::Env) {
let traces = env.trace_store().read(&a.run).unwrap_or_else(|e| {
eprintln!("aura: reading run '{}' traces failed: {e}", a.run);
std::process::exit(1);
});
let series = |tap: &str| -> Vec<(Timestamp, f64)> {
let ct = traces.taps.iter().find(|t| t.tap == tap).unwrap_or_else(|| {
let have: Vec<&str> = traces.taps.iter().map(|t| t.tap.as_str()).collect();
eprintln!("aura: run '{}' has no tap '{tap}' (taps: {have:?})", a.run);
std::process::exit(1);
});
// column 0 is the tap's value series; `f64_field` (already imported in main.rs,
// used by run_signal_r) projects field 0 of the (Timestamp, Vec<Scalar>) rows
// ColumnarTrace::to_rows rebuilds (all cells are Scalar::f64, report.rs:248, so
// f64_field never hits its kind-mismatch panic).
f64_field(&ct.to_rows(), 0)
};
let signal = series(&a.signal);
let price = series(&a.price);
let out = information_coefficient(&signal, &price, a.horizon, a.permutations, a.seed);
let report = IcReport {
run: a.run,
signal_tap: a.signal,
price_tap: a.price,
horizon: a.horizon,
permutations: a.permutations,
seed: a.seed,
n_pairs: out.n_pairs,
information_coefficient: out.information_coefficient,
overfit_probability: out.overfit_probability,
};
println!("{}", report.to_json());
}
/// `aura data list` (#264 cut 2): print the archive's known symbols, sorted
/// ascending, one per line — the discovery step before `aura data coverage`
/// or scoping a campaign's instrument matrix. Resolves the archive root
@@ -4498,6 +4651,7 @@ fn main() {
Command::Process(a) => research_docs::process_cmd(a, &env),
Command::Campaign(a) => research_docs::campaign_cmd(a, &env),
Command::Data(a) => dispatch_data(a, &env),
Command::Measure(a) => dispatch_measure(a, &env),
}
}
@@ -6526,3 +6680,114 @@ mod tests {
assert_eq!(stop_k, R_SMA_STOP_K, "omitted --stop-k defaults to the regime");
}
}
#[cfg(test)]
mod ic_tests {
use super::*;
use aura_core::Timestamp;
fn ts(i: i64) -> Timestamp {
Timestamp(i) // Timestamp is a tuple struct over epoch-ns i64 (pub field, per report.rs)
}
#[test]
fn ic_engineered_signal_is_significant() {
// price path; forward return fr[i] = (p[i+1]-p[i])/p[i]. Engineer signal_t = fr[i]
// exactly (a hand-built OFFLINE series — a look-ahead signal is impossible in a
// causal run, C2, so this property lives at the unit level, not E2E).
let price: Vec<(Timestamp, f64)> = [100.0, 101.0, 100.5, 102.0, 101.0, 103.0, 102.5, 104.0]
.iter()
.enumerate()
.map(|(i, &p)| (ts(i as i64), p))
.collect();
let signal: Vec<(Timestamp, f64)> = (0..price.len() - 1)
.map(|i| (ts(i as i64), (price[i + 1].1 - price[i].1) / price[i].1))
.collect();
let out = information_coefficient(&signal, &price, 1, 1000, 0);
assert!(out.information_coefficient > 0.99, "ic = {}", out.information_coefficient);
assert!(out.overfit_probability < 0.05, "p = {}", out.overfit_probability);
assert!(out.n_pairs >= 6);
}
#[test]
fn ic_constant_signal_is_not_significant() {
let price: Vec<(Timestamp, f64)> = [100.0, 101.0, 100.5, 102.0, 101.0, 103.0]
.iter().enumerate().map(|(i, &p)| (ts(i as i64), p)).collect();
let signal: Vec<(Timestamp, f64)> =
(0..price.len()).map(|i| (ts(i as i64), 1.0)).collect(); // constant → zero variance
let out = information_coefficient(&signal, &price, 1, 1000, 0);
assert_eq!(out.information_coefficient, 0.0);
assert_eq!(out.overfit_probability, 1.0);
}
#[test]
fn ic_varying_uncorrelated_signal_is_not_significant() {
// The false-positive control on a NON-degenerate null: a VARYING signal (unlike the
// constant case above, whose null is degenerate) that is exactly uncorrelated with the
// forward returns. price is chosen so forward returns are exactly [0.01, 0.02, 0.03,
// 0.04]; signal is a permutation of those values arranged orthogonal to them
// (Σ centred products = 0 → IC = 0), so the permutation null genuinely varies yet the
// raw IC sits in its bulk (~half the permutations exceed it), not its tail → not
// significant. This is the core guarantee of a significance test.
let price: Vec<(Timestamp, f64)> = [100.0, 101.0, 103.02, 106.1106, 110.355024]
.iter().enumerate().map(|(i, &p)| (ts(i as i64), p)).collect();
let signal: Vec<(Timestamp, f64)> = [0.03, 0.01, 0.04, 0.02]
.iter().enumerate().map(|(i, &s)| (ts(i as i64), s)).collect();
let out = information_coefficient(&signal, &price, 1, 1000, 0);
assert_eq!(out.n_pairs, 4);
assert!(out.information_coefficient.abs() < 1e-6, "ic ~ 0 expected, got {}", out.information_coefficient);
assert!(
out.overfit_probability > 0.1,
"an uncorrelated signal must not be significant: p = {}",
out.overfit_probability
);
}
#[test]
fn ic_is_deterministic_given_seed() {
let price: Vec<(Timestamp, f64)> = [100.0, 101.0, 100.5, 102.0, 101.0, 103.0, 102.5]
.iter().enumerate().map(|(i, &p)| (ts(i as i64), p)).collect();
let signal: Vec<(Timestamp, f64)> =
[0.3, -0.1, 0.4, -0.2, 0.1, 0.5, -0.3].iter().enumerate()
.map(|(i, &s)| (ts(i as i64), s)).collect();
let a = information_coefficient(&signal, &price, 1, 500, 7);
let b = information_coefficient(&signal, &price, 1, 500, 7);
assert_eq!(a.information_coefficient, b.information_coefficient);
assert_eq!(a.overfit_probability, b.overfit_probability);
assert_eq!(a.n_pairs, b.n_pairs);
}
#[test]
fn ic_degenerate_floor_on_too_few_pairs() {
let price = vec![(ts(0), 100.0)]; // one row → no forward return → 0 pairs
let signal = vec![(ts(0), 1.0)];
let out = information_coefficient(&signal, &price, 1, 1000, 0);
assert_eq!(out.information_coefficient, 0.0);
assert_eq!(out.overfit_probability, 1.0);
assert_eq!(out.n_pairs, 0);
}
#[test]
fn ic_pairs_only_on_exact_timestamp_overlap() {
// Property: alignment is by EXACT ts match, unmatched dropped (the cross-cadence
// case). A price ts carrying no signal is dropped, and a signal ts absent from the
// price spine is ignored — so n_pairs is the OVERLAP count, never the price length,
// and the dropped rows never enter the correlation.
//
// price spine ts 0..8 → horizon-1 forward returns exist at ts 0..7. Signal is placed
// ONLY at ts {1,2,4,5} (leaving price ts {0,3,6} without a signal → dropped) plus two
// decoy ts {50,60} absent from the price spine (→ never looked up). Where present, the
// signal equals the forward return exactly, so the 4 matched pairs correlate perfectly.
let price: Vec<(Timestamp, f64)> =
[100.0, 101.0, 100.5, 102.0, 101.0, 103.0, 102.5, 104.0]
.iter().enumerate().map(|(i, &p)| (ts(i as i64), p)).collect();
let fr = |i: usize| (price[i + 1].1 - price[i].1) / price[i].1;
let mut signal: Vec<(Timestamp, f64)> =
[1usize, 2, 4, 5].iter().map(|&i| (ts(i as i64), fr(i))).collect();
signal.push((ts(50), 999.0)); // decoy ts not on the price spine → unmatched, dropped
signal.push((ts(60), -999.0));
let out = information_coefficient(&signal, &price, 1, 1000, 0);
assert_eq!(out.n_pairs, 4, "only the 4 overlapping ts pair; price ts 0/3/6 and the decoys drop");
assert!(out.information_coefficient > 0.99, "ic = {}", out.information_coefficient);
}
}