neuroalpha 0.2.0 → 0.3.12

package/server/model/forecast.js

@@ -1,3 +1,4 @@
+// Copyright (c) 2026 Roman A. Perepelkin. Licensed under the MIT License.
 import { normalizeParams } from '../settings.js';
 import { round, DAY, fmtBarMsk } from '../util.js';
 import { loadDataset, barsFromHours, BAR_MS, BAR_INTERVAL } from './dataset.js';
@@ -18,37 +19,28 @@ export const lastClosedIndex = (ds, now = Date.now()) => {
 // Обучение на окне, заканчивающемся свечой idx, и прогноз по этой свече.
 // Возвращает строки по всем акциям: ожидаемое движение (относительно рынка), сигнал покупки, кто влияет.
 export async function predictAt(ds, p, idx, { explain = true } = {}) {
-  const H = barsFromHours(p.modelHorizon);
+  const autoAdapt = p.modelAutoAdapt !== false;
+  const effHorizon = autoAdapt ? 2.5 : p.modelHorizon;
+  const effMinEdge = autoAdapt ? 2.5 : p.modelMinEdge;
+  const H = barsFromHours(effHorizon);
+  const maxH = autoAdapt ? Math.round(H * 1.4) : H;
   const trainLen = Math.round(p.modelTrainDays * ds.barsPerDay);
-  const trainStart = Math.max(FEATURE_WARMUP + H + 2, idx - trainLen);
+  const trainStart = Math.max(FEATURE_WARMUP + maxH + 2, idx - trainLen);
   // «текстовая» часть карты: устойчивые описания бизнеса (профили готовятся отдельно, здесь только читаем)
   const profiles = p.modelTextWeight > 0 ? loadProfiles() : null;
-  const f = buildFold(ds, { trainStart, testStart: idx, testEnd: idx + 1, H, embedDim: p.modelEmbedDim, textWeight: p.modelTextWeight, profiles });
+  const f = buildFold(ds, { trainStart, testStart: idx, testEnd: idx + 1, H, embedDim: p.modelEmbedDim, textWeight: p.modelTextWeight, profiles, multiHorizon: autoAdapt });
   const u = idx - f.lo;
   const chosen = p.modelType in MODELS ? p.modelType : 'attention';
-  const need = ['attention', chosen];
-  // трансформер обучается здесь, только если торгует он (для сравнения — только в исследовании)
-  const list = [...new Set(explain ? [...Object.keys(MODELS).filter((n) => !MODELS[n].heavy), chosen] : need)];
+  const list = [...new Set(explain ? Object.keys(MODELS) : ['attention', chosen])];
   const fitted = {};
   const preds = {};
   for (const n of list) {
-    if (n === 'transformer') {
-      // ~минута обучения — в отдельном потоке (или на видеокарте), чтобы сервер не замирал
-      const tm = await import('./transformer.js');
-      const { fitPredictIsolated } = await import('./transformer-pool.js');
-      const r = await fitPredictIsolated(tm.transformerInput(f), [u], tm.transformerCfg(p), { gpu: p.transformerGpu });
-      fitted[n] = { name: n, info: r.info, where: r.where };
-      preds[n] = r.preds.get(u);
-      continue;
-    }
     fitted[n] = await MODELS[n].fit(f);
     preds[n] = MODELS[n].predict(fitted[n], f, u);
-    if (fitted[n].dispose) fitted[n].dispose();
   }
   const cost = (p.commissionPct + p.slippagePct) / 100;
-  const threshold = p.modelMinEdge * 2 * cost;
+  const threshold = effMinEdge * 2 * cost;
   const bear = isMarketDown(ds, idx);
-  const autoAdapt = p.modelAutoAdapt !== false;
   const dir = p.modelDirection || 'long';
   const needL = autoAdapt && bear && dir !== 'neutral' ? threshold * 1.25 : threshold;
   const needS = autoAdapt && dir === 'longshort' ? (bear ? threshold * 0.90 : threshold * 1.25) : threshold;
@@ -94,7 +86,10 @@ function withNews(result, p) {
 // Прогноз на ближайшие H часов по последней закрытой свече
 export async function forecastNow(input, { force = false, onProgress } = {}) {
   const p = normalizeParams(input);
-  const key = JSON.stringify([p.modelType, p.modelHorizon, p.modelTrainDays, p.modelEmbedDim, p.modelMinEdge, p.modelDirection, p.modelUseMacro, p.modelUseDividends, p.modelTextWeight, p.commissionPct, p.slippagePct, ...(p.modelType === 'transformer' ? [p.transformerEpochs, p.transformerStride, p.transformerGpu] : [])]);
+  const autoAdapt = p.modelAutoAdapt !== false;
+  const effHorizon = autoAdapt ? 2.5 : p.modelHorizon;
+  const effMinEdge = autoAdapt ? 2.5 : p.modelMinEdge;
+  const key = JSON.stringify([autoAdapt, p.modelType, effHorizon, p.modelTrainDays, p.modelEmbedDim, effMinEdge, p.modelDirection, p.modelUseMacro, p.modelUseDividends, p.modelTextWeight, p.commissionPct, p.slippagePct]);
   if (!force && cache && cache.key === key && Date.now() - cache.at < 10 * 60000) return withNews(cache.result, p);
   if (!force && inflight.has(key)) return withNews(await inflight.get(key), p);
   const task = (async () => {
@@ -107,11 +102,10 @@ export async function forecastNow(input, { force = false, onProgress } = {}) {
       at: now,
       barTs: ds.timeline[last],
       bar: fmtBarMsk(ds.timeline[last]),
-      horizon: p.modelHorizon, direction: p.modelDirection, model: r.chosen, thresholdPct: round(r.threshold * 100, 3), marketDown: r.marketDown,
+      horizon: effHorizon, direction: p.modelDirection, model: r.chosen, thresholdPct: round(r.threshold * 100, 3), marketDown: r.marketDown,
       rows: r.rows.map(({ i, expected, ...x }) => ({ ...x, cluster: snap.clusters[i] })).sort((a, b) => b.expectedPct - a.expectedPct),
       map: { symbols: ds.tokens.map((t) => t.symbol), names: ds.tokens.map((t) => t.name), pts: snap.pts.map(([x, y]) => [round(x, 3), round(y, 3)]), cl: snap.clusters },
       heads: HEADS.map((h, i) => ({ label: h.label, coef: round(r.fitted.attention.coef[i], 4) })),
-      ...(r.fitted.transformer ? { transformer: { device: r.fitted.transformer.where, bestEpoch: r.fitted.transformer.info.bestEpoch, epochs: r.fitted.transformer.info.epochs } } : {}),
     };
     cache = { key, at: now, result };
     return result;

package/server/model/linalg.js

@@ -1,3 +1,4 @@
+// Copyright (c) 2026 Roman A. Perepelkin. Licensed under the MIT License.
 // Небольшая линейная алгебра без зависимостей: матрицы — массивы Float64Array (строки)
 
 export const matrix = (n, m) => Array.from({ length: n }, () => new Float64Array(m));