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>
This commit is contained in:
Ovidiu U
2026-05-03 08:40:05 +01:00
parent d13a29df01
commit ddd591ad47
63 changed files with 5109 additions and 13 deletions

View File

@@ -0,0 +1,209 @@
<?php
namespace App\Services\Forecasting;
use App\Models\BrentPrice;
use App\Models\LlmOverlay;
use App\Models\VolatilityRegime;
use App\Models\WatchedEvent;
use Illuminate\Support\Facades\DB;
/**
* Layer 5 sole owner of `volatility_regimes.active`. Hourly cron.
*
* OR-combines four triggers:
* 1. Brent close-to-close move > 3% (FRED `DCOILBRENTEU`).
* 2. Most recent `llm_overlays.major_impact_event = true` AND at
* least one verified URL.
* 3. `station_prices` daily churn > 1.5× 30-day baseline. Gated
* until 180 days of polling toggleable via config.
* 4. `watched_events` row covering today.
*
* When the flag flips ON, an event-driven LLM refresh is queued
* (Layer 4 enforces its own 4h cooldown). When OFF, the row is
* closed with `flipped_off_at`.
*/
final class VolatilityRegimeService
{
private const float BRENT_MOVE_PCT = 3.0;
private const float STATION_CHURN_RATIO = 1.5;
private const int STATION_CHURN_MIN_POLLING_DAYS = 180;
public function __construct(
private readonly LlmOverlayService $llmOverlay,
) {}
public function evaluate(): ?VolatilityRegime
{
$trigger = $this->detectTrigger();
$current = VolatilityRegime::currentlyActive();
if ($trigger !== null && $current === null) {
$row = $this->flipOn($trigger);
$this->llmOverlay->run(eventDriven: true);
return $row;
}
if ($trigger === null && $current !== null) {
$this->flipOff($current);
return null;
}
return $current;
}
/** @return array{type: string, detail: string}|null */
private function detectTrigger(): ?array
{
return $this->brentMoveTrigger()
?? $this->llmEventTrigger()
?? $this->stationChurnTrigger()
?? $this->watchedEventTrigger();
}
/** @return array{type: string, detail: string}|null */
private function brentMoveTrigger(): ?array
{
$rows = BrentPrice::query()
->orderByDesc('date')
->limit(2)
->get(['date', 'price_usd']);
if ($rows->count() < 2) {
return null;
}
$latest = (float) $rows[0]->price_usd;
$prior = (float) $rows[1]->price_usd;
if ($prior === 0.0) {
return null;
}
$pctMove = abs(($latest - $prior) / $prior) * 100;
if ($pctMove <= self::BRENT_MOVE_PCT) {
return null;
}
$direction = $latest > $prior ? '+' : '-';
return [
'type' => 'brent_move',
'detail' => sprintf('Brent %s%.2f%% (%s → %s)', $direction, $pctMove, $rows[1]->date->toDateString(), $rows[0]->date->toDateString()),
];
}
/** @return array{type: string, detail: string}|null */
private function llmEventTrigger(): ?array
{
$latest = LlmOverlay::query()->orderByDesc('ran_at')->first();
if ($latest === null || ! $latest->major_impact_event) {
return null;
}
$hasVerifiedUrl = collect((array) $latest->events_json)
->contains(fn ($e): bool => is_array($e) && ! empty($e['url']));
if (! $hasVerifiedUrl) {
return null;
}
$headline = collect((array) $latest->events_json)->pluck('headline')->filter()->first();
return [
'type' => 'llm_event',
'detail' => sprintf('LLM major impact: %s', $headline ?? 'unspecified'),
];
}
/** @return array{type: string, detail: string}|null */
private function stationChurnTrigger(): ?array
{
if (! $this->stationChurnEnabled()) {
return null;
}
$oldest = DB::table('station_prices')->min('price_effective_at');
if ($oldest === null) {
return null;
}
$pollingDays = (int) abs(now()->diffInDays($oldest));
if ($pollingDays < self::STATION_CHURN_MIN_POLLING_DAYS) {
return null;
}
$last24h = (int) DB::table('station_prices')
->where('price_effective_at', '>=', now()->subDay())
->distinct('station_id')
->count('station_id');
$baseline = (int) DB::table('station_prices')
->where('price_effective_at', '>=', now()->subDays(30))
->where('price_effective_at', '<', now()->subDay())
->distinct('station_id')
->count('station_id');
if ($baseline === 0) {
return null;
}
$dailyBaseline = $baseline / 29; // 29 days of history before yesterday
if ($last24h <= $dailyBaseline * self::STATION_CHURN_RATIO) {
return null;
}
return [
'type' => 'station_churn',
'detail' => sprintf('Station churn %d/24h vs %.1f baseline (%.2fx)', $last24h, $dailyBaseline, $last24h / $dailyBaseline),
];
}
/** @return array{type: string, detail: string}|null */
private function watchedEventTrigger(): ?array
{
$row = WatchedEvent::query()
->where('starts_at', '<=', now())
->where('ends_at', '>=', now())
->orderBy('starts_at')
->first();
if ($row === null) {
return null;
}
return [
'type' => 'manual',
'detail' => sprintf('Watched event: %s', $row->label),
];
}
private function stationChurnEnabled(): bool
{
return (bool) config('services.forecasting.station_churn_enabled', false);
}
/** @param array{type: string, detail: string} $trigger */
private function flipOn(array $trigger): VolatilityRegime
{
return VolatilityRegime::query()->create([
'flipped_on_at' => now(),
'flipped_off_at' => null,
'trigger' => $trigger['type'],
'trigger_detail' => $trigger['detail'],
'active' => true,
]);
}
private function flipOff(VolatilityRegime $row): void
{
$row->update([
'flipped_off_at' => now(),
'active' => false,
]);
}
}