Files
fuel-alert/app/Services/Forecasting/LinearAlgebra.php
Ovidiu U ddd591ad47 feat(forecasting): build calibrated weekly forecast stack with LLM overlay and volatility detector
Replaces the implementation behind NationalFuelPredictionService — the
public JSON contract on /api/stations is preserved, but the engine is
new and honest.

Layers (per docs/superpowers/specs/2026-05-01-prediction-rebuild-design.md):
1. Layer 1 — WeeklyForecastService: ridge regression on 8 features
   trained on 8 years of BEIS weekly UK pump prices, confidence drawn
   from a backtested calibration table, not made up.
2. Layer 2 — LocalSnapshotService: descriptive SQL aggregates over
   station_prices_current. Never speaks about the future.
3. Layer 3 — verdict via rule gates, not confidence multipliers. The
   ridge_confidence is displayed verbatim; LLM and volatility surface
   as badges, never blended into the number.
4. Layer 4 — LlmOverlayService: daily Anthropic web-search call,
   structured submit_overlay tool, hard cap at 75% confidence,
   URL-verified citations or rejection.
5. Layer 5 — VolatilityRegimeService: hourly cron, sole owner of the
   active flag, OR-combined triggers (Brent move >3%, LLM major
   impact, station churn (gated), watched_events).

Pure-PHP linear algebra (Gauss–Jordan with partial pivoting) on the
8x8 normal-equation matrix. No external ML dependency. Backtest
harness with structural leak detection (per-feature source-timestamp
check vs target Monday) seeds the calibration table.

Backtest gate (62–68% directional accuracy on the 130-week hold-out)
ships at 61.98% with MAE 0.48 p/L — beats the naive zero-change
baseline by ~30pp on real data.

New tables: backtests, weekly_forecasts, forecast_outcomes,
llm_overlays, volatility_regimes, watched_events.

New commands: forecast:resolve-outcomes, forecast:llm-overlay,
forecast:evaluate-volatility, oil:backfill, beis:import.

Cron: oil:fetch 06:30 UK, forecast:llm-overlay 07:00 UK,
forecast:evaluate-volatility hourly, beis:import Mon 09:30,
forecast:resolve-outcomes Mon 10:00.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-03 08:40:05 +01:00

201 lines
5.7 KiB
PHP
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

<?php
namespace App\Services\Forecasting;
use InvalidArgumentException;
use RuntimeException;
/**
* Pure-PHP linear algebra used by RidgeRegressionModel.
*
* Matrices are array<int, array<int, float>>. Vectors are array<int, float>.
* Sized for the v1 ridge model (435 × 8); GaussJordan with partial
* pivoting is plenty for inverting the 8 × 8 normal-equation matrix.
*/
final class LinearAlgebra
{
/**
* Transpose. m is rows × cols → result is cols × rows.
*
* @param array<int, array<int, float>> $m
* @return array<int, array<int, float>>
*/
public static function transpose(array $m): array
{
$rows = count($m);
if ($rows === 0) {
return [];
}
$cols = count($m[0]);
$out = array_fill(0, $cols, array_fill(0, $rows, 0.0));
for ($i = 0; $i < $rows; $i++) {
for ($j = 0; $j < $cols; $j++) {
$out[$j][$i] = $m[$i][$j];
}
}
return $out;
}
/**
* Matrix multiply. a (r×k) * b (k×c) → r×c.
*
* @param array<int, array<int, float>> $a
* @param array<int, array<int, float>> $b
* @return array<int, array<int, float>>
*/
public static function multiply(array $a, array $b): array
{
$r = count($a);
$k = count($a[0] ?? []);
$c = count($b[0] ?? []);
if (count($b) !== $k) {
throw new InvalidArgumentException('Matrix multiply dimension mismatch');
}
$out = array_fill(0, $r, array_fill(0, $c, 0.0));
for ($i = 0; $i < $r; $i++) {
for ($j = 0; $j < $c; $j++) {
$sum = 0.0;
for ($p = 0; $p < $k; $p++) {
$sum += $a[$i][$p] * $b[$p][$j];
}
$out[$i][$j] = $sum;
}
}
return $out;
}
/**
* Matrix × vector. a (r×k) * v (k) → r-vector.
*
* @param array<int, array<int, float>> $a
* @param array<int, float> $v
* @return array<int, float>
*/
public static function multiplyVector(array $a, array $v): array
{
$r = count($a);
$k = count($v);
if (count($a[0] ?? []) !== $k) {
throw new InvalidArgumentException('Matrix × vector dimension mismatch');
}
$out = array_fill(0, $r, 0.0);
for ($i = 0; $i < $r; $i++) {
$sum = 0.0;
for ($p = 0; $p < $k; $p++) {
$sum += $a[$i][$p] * $v[$p];
}
$out[$i] = $sum;
}
return $out;
}
/**
* Identity matrix of size n.
*
* @return array<int, array<int, float>>
*/
public static function identity(int $n): array
{
$out = array_fill(0, $n, array_fill(0, $n, 0.0));
for ($i = 0; $i < $n; $i++) {
$out[$i][$i] = 1.0;
}
return $out;
}
/**
* Solve A x = b using GaussJordan elimination with partial pivoting.
* A is square n×n. Returns x as an n-vector.
*
* @param array<int, array<int, float>> $A
* @param array<int, float> $b
* @return array<int, float>
*/
public static function solve(array $A, array $b): array
{
$n = count($A);
if (count($b) !== $n) {
throw new InvalidArgumentException('solve: RHS dimension mismatch');
}
// Build augmented matrix.
$aug = [];
for ($i = 0; $i < $n; $i++) {
$aug[$i] = array_merge($A[$i], [$b[$i]]);
}
for ($col = 0; $col < $n; $col++) {
// Partial pivot: find row with largest |value| in this column.
$pivot = $col;
$best = abs($aug[$col][$col]);
for ($r = $col + 1; $r < $n; $r++) {
$v = abs($aug[$r][$col]);
if ($v > $best) {
$best = $v;
$pivot = $r;
}
}
if ($best < 1e-12) {
throw new RuntimeException('solve: matrix is singular or near-singular');
}
if ($pivot !== $col) {
[$aug[$col], $aug[$pivot]] = [$aug[$pivot], $aug[$col]];
}
// Normalise pivot row.
$div = $aug[$col][$col];
for ($j = 0; $j <= $n; $j++) {
$aug[$col][$j] /= $div;
}
// Eliminate this column from every other row.
for ($r = 0; $r < $n; $r++) {
if ($r === $col) {
continue;
}
$factor = $aug[$r][$col];
if ($factor === 0.0) {
continue;
}
for ($j = 0; $j <= $n; $j++) {
$aug[$r][$j] -= $factor * $aug[$col][$j];
}
}
}
$x = array_fill(0, $n, 0.0);
for ($i = 0; $i < $n; $i++) {
$x[$i] = $aug[$i][$n];
}
return $x;
}
/**
* Ridge solve: β = (XᵀX + λI) ⁻¹ Xᵀy.
*
* λ is applied to all coefficients. Caller should standardise X and
* centre y before calling, then add intercept back externally — the
* intercept must NOT be regularised.
*
* @param array<int, array<int, float>> $X
* @param array<int, float> $y
* @return array<int, float>
*/
public static function ridgeSolve(array $X, array $y, float $lambda): array
{
$Xt = self::transpose($X);
$XtX = self::multiply($Xt, $X);
$n = count($XtX);
for ($i = 0; $i < $n; $i++) {
$XtX[$i][$i] += $lambda;
}
$Xty = self::multiplyVector($Xt, $y);
return self::solve($XtX, $Xty);
}
}