suggest_patterns_from_market_context() was missing the specialist desk
block that score_patterns_with_context() already received. All 7 desks
(forex, metals, agri, energy, indices, crypto, bonds) with their
fundamentals, macro sensitivity, and upcoming reports are now injected
so the AI can generate targeted patterns per desk rather than generic ones.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- database.py: add trade_budget_eur / preferred_horizon_min/max config
defaults and include them in cycle config migrations
- auto_cycle.py: read trade params from config and inject into cycle_meta
- ai_analyzer.py: inject INVESTOR TRADE MANDATE block into scoring and
suggestion prompts so GPT-4o penalises horizon mismatches and sizes
within the capital cap
- Config.tsx: Trade Parameters card with budget + horizon sliders and live
mandate summary
- TradeIdeas.tsx: horizon filter pills (< 1M / 1-3M / 3-6M / > 6M) and
budget/horizon indicator pulled from saved config
- useApi.ts: extend useUpdateCycleConfig type with new config fields
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- database.py: add timedelta to datetime import (used in
get_recent_economic_surprises, was raising NameError)
- ai_analyzer.py: scoring split was searching French string
'Retourne UNIQUEMENT ce JSON valide:' but prompt is now in English
'Return ONLY this valid JSON:' — caused IndexError crashing every cycle
- requirements.txt: add lxml>=5.0.0 (yfinance earnings_dates dependency,
was silently failing all 23 ticker fetches every hour)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- New institutional_reports table (DB) with importance, signals per asset class, key points, absorption tracking
- cot_fetcher.py: CFTC Socrata API (6dca-aqww), 7 instruments (Gold/Silver/Copper/WTI/NatGas/SP500/EURUSD), net positioning + 52-week z-score
- eia_fetcher.py: EIA API v2, 4 series (crude/Cushing/gasoline/distillates), WoW surprise detection
- institutional.py router: GET /reports, GET /reports/{id}, POST /refresh, GET /stats
- institutional_scheduler.py: weekly auto-fetch (COT Saturdays, EIA Wednesday afternoons)
- ai_analyzer.py: build_institutional_block() + institutional_block param injected into AI scoring prompt
- auto_cycle.py: inject institutional block into suggestion + scoring, absorption tracking via keyword overlap after each cycle commentary
- InstitutionalReports.tsx: full page with filter bar (type/category/importance/period), cards with key point bullets, EXTREME alerts highlighted, signal badges, absorption badge, trading implications, expandable detail
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Complete French→English translation across all frontend pages and backend
services — every label, button, header, empty state, toast, and nav item
is now in English. Build verified clean (tsc + vite). No i18n library
added; direct string replacement throughout.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Phase 1 — Catégorisation:
- database.py: ADD COLUMN category + signal_direction on custom_patterns (migration);
save_custom_pattern persists category/signal_direction; new helpers:
get_unclassified_patterns(), update_pattern_classification(),
get_patterns_with_last_score()
- ai_analyzer.py: PATTERN_CATEGORIES dict (8 categories: géopolitique, macro_monétaire,
technique, commodités_supply, risk_off, flux_saisonnier, géo_économique, crédit_stress);
classify_patterns_batch() → GPT-4o-mini batch classification
- suggest schema: added category + signal_direction fields so new patterns are
classified from birth
- auto_cycle.py: Step 3.1 classifies all unclassified patterns after each suggestion
Phase 2 — Convergence layer (post-scoring, no extra AI call):
- ai_analyzer.py: _compute_convergence() groups scored patterns by (underlying, signal_direction);
conviction_bonus = min(20, +5 per additional agreeing pattern); adds conviction_score,
conviction_bonus, convergence_count, convergence_underlying, convergence_partners to each result;
called at end of score_patterns_with_context(), re-sorts by conviction_score
- auto_cycle.py: logs convergence summary after scoring; propagates category/signal_direction
to scored results for display
Phase optionnelle — Convergence in suggestion prompt:
- ai_analyzer.py: suggest_patterns_from_market_context() accepts convergence_block param;
injected into prompt so AI knows which underlyings have multi-pattern agreement
- auto_cycle.py: before suggestion, loads last-cycle scores via get_patterns_with_last_score(),
calls _compute_convergence() to build convergence block, passes to suggester
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- auto_cycle.py: detect weekend/market session, build cycle_meta with day_of_week/is_weekend/market_note;
IVGate skips iv_rank>=99 on weekends to avoid artificial weekend option premium cascade;
inject portfolio context (open trades + price moves + concentration) before AI scoring;
pass portfolio_context_block + run_id to both AI scorer and suggester
- ai_analyzer.py: _build_temporal_news_block injects market session banner (WEEKEND warning,
pre/after-market note, or open session label) so AI knows markets are closed and defers execution to Monday
- iv_engine.py: add WHEAT/EUR/USD ticker aliases; skip saving IV snapshots on weekends to protect history;
resolve aliases before slash-format conversion in _resolve_ticker
- technical_indicators.py: fix pandas MultiIndex from yfinance>=0.2 (droplevel+squeeze);
use period proportional to lookback instead of fixed period=1d
- database.py: asset_class ticker-based fallback (_asset_class_from_ticker); one-time backfill migration
for all NULL asset_class rows; ai_call_logs table + save/get helpers; normalize_ticker public function
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Portfolio context (portfolio_context.py):
- get_open_trades_with_moves(): fetches open trades + 1d/5d yfinance price moves
- get_portfolio_concentration(): counts by asset_class
- build_portfolio_context_block(): formatted prompt block with strict AI instructions
(no double positions, flag contradictions, avoid overweight classes)
AI call logging:
- ai_call_logs table in DB (run_id, call_type, system/user prompt, response, tokens, ms)
- _chat() now accepts log_meta dict → saves call to DB non-blocking after each call
- suggest and score_batch calls pass run_id + call_type for full traceability
auto_cycle.py:
- Builds portfolio context before snapshot and both AI calls
- Context snapshot now includes portfolio_open_positions key
SystemLogs.tsx:
- "Contexte IA" tab gains sub-tabs: Contexte / Appels IA
- AiCallRow: expandable with 3 panes (user prompt / system prompt / response)
shows model, tokens breakdown, duration, call type badge
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Direction (Ouvert) : t.direction n'existe pas dans trade_entry_prices → utilisait
undefined, excluait tout. Remplacé par _isBearishStr(t.strategy) comme Fermés.
Asset class (Ouvert, Fermés, Non loggés) : l'IA retournait parfois "commodities",
"currencies", "fx", "equity" au lieu des clés canoniques. Double correction :
- Frontend : _normalizeAssetClass() mappe les variantes → energy|metals|agriculture|
indices|equities|forex dans les 3 sections filtrées
- Backend database.py : _normalize_asset_class() appliqué à l'INSERT dans
trade_entry_prices et skipped_trades (nouveaux trades normalisés au stockage)
- Prompt ai_analyzer.py : suggested_trades[].asset_class contraint à l'enum explicite
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Phase 4 — Price Discovery Status (la pièce maîtresse) :
- price_discovery.py (nouveau) : capture_price_snapshots() sauve les prix des tickers
liés à chaque news scorée (energy→BZ=F/NG=F, metals→GC=F/HG=F, indices→^GSPC/IWM)
- compute_absorptions() mesure combien du mouvement attendu s'est déjà produit
(status: not_yet_priced <30% / partially_priced 30-80% / fully_priced >80%)
- build_price_discovery_block() → bloc prompt avec opportunités classées
- database.py : table news_price_snapshots + save/get/purge fonctions
- auto_cycle.py : capture après ai_score_news_batch, compute avant suggestions,
block injecté dans suggestion + scoring prompts + context snapshot
- ai_analyzer.py : param price_discovery_block dans suggest + score
Phase 5 — Replay historique :
- cycle.py : POST /api/cycle/contexts/{run_id}/replay — recharge le snapshot historique
et relance suggest_patterns_from_market_context avec le contexte original
- useApi.ts : hook useReplayCycle
- SystemLogs.tsx : bouton "Rejouer ce cycle" dans onglet Contexte IA avec champ
notes, résultats inline (liste des patterns générés), section price_discovery
ouverte par défaut en rouge
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Phase 2 — Données macro FRED :
- fred_fetcher.py (nouveau) : 7 séries FRED (CPI, NFP, UNRATE, FEDFUNDS, GDP, ICSA,
spread 10Y-2Y) avec détection direction bullish/bearish et block prompt formaté
- ai_analyzer.py : param fred_block dans suggest + score, injecté dans les deux prompts
- auto_cycle.py : fetch FRED non-bloquant avant la suggestion
Context log — Snapshot du contexte complet :
- database.py : table cycle_context_snapshots + save/get/list fonctions
- auto_cycle.py : sauvegarde le snapshot (meta, news partitionnées, FRED, tech, IV, quotes)
- cycle.py : GET /api/cycle/contexts + GET /api/cycle/contexts/{run_id}
- useApi.ts : hooks useCycleContextSnapshots + useCycleContextSnapshot
- SystemLogs.tsx : onglet "Contexte IA" avec liste de cycles et visualiseur JSON
par section (cycle_meta, macro, news, FRED, tech) avec accordéon
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- database.py: get_last_completed_cycle_ts() pour mesurer le delta entre cycles
- auto_cycle.py: calcul delta_minutes + cycle_meta dict transmis aux fonctions IA
- ai_analyzer.py: apply_news_decay() (halflife par catégorie), partition_news_by_age()
(3 buckets: inter_cycle / recent_24h / older), _build_temporal_news_block() pour
le prompt suggestion; cycle_meta injecté aussi dans score_patterns_with_context()
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- auto_cycle.py: replace _log with logger (NameError at lines 484/489)
- auto_cycle.py: normalize underlying via _normalize_ticker before _resolve_ticker
so WHEAT→ZW=F→WEAT and EUR/USD→EURUSD=X→FXE reach the IV watchlist correctly
- iv_engine.py: _resolve_ticker now strips slash-format forex (EUR/USD→EURUSD=X)
before _PROXY lookup, fixing yfinance 500/404 spam from get_atm_iv
- database.py: _fetch in log_trade_entries uses _normalize_ticker (not _normalize_yf_ticker)
so commodity aliases like WHEAT→ZW=F are applied at price-fetch time
- ai_analyzer.py: max_workers=1 for batch scorer — parallel workers both slept and
retried simultaneously after 429, causing repeated bursts; sequential fixes the pattern
- journal.py + JournalDeBord.tsx: add price_warning field (no_price_data/no_entry_price/
no_live_price) with visible ⚠ badge and amber color on affected ticker/price cells
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Ticker normalization (_normalize_ticker):
- EUR/USD slash-format → EURUSD=X (was passed raw to yfinance → 500/404 spam)
- bare 6-char forex pairs EURUSD/USDJPY etc → append =X
- commodity alias table: WHEAT→ZW=F, CORN→ZC=F, WTI→CL=F, BRENT→BZ=F,
GOLD→GC=F, SILVER→SI=F, NATGAS→NG=F, SUGAR→SB=F, + 15 others
- also normalize underlying at log_trade_entries time so stored tickers
are already canonical before MtM lookups
GPT-4o 429 rate limit:
- _chat() retries up to 3× on rate_limit errors, respects retry-after hint
from error message (e.g. "try again in 12.37s"), falls back to 2^n×5s
- batch scorer: parallel workers 4→2 to halve the token burst per cycle
(2 concurrent batches × ~6K tokens vs 4 × ~6K = 24K burst at 30K limit)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Sprint 2.1 — Pattern Reliability Score
- database.py: get_pattern_reliability() — win_rate × log(n+1) sur trades matures (≥35% horizon)
- database.py: get_all_pattern_reliability_map() pour injection rapide dans les prompts
- ai_analyzer.py: inject reliability_map dans suggest_patterns (patterns fiables mis en avant)
- auto_cycle.py: charge reliability_map avant suggestion et le passe au suggéreur
- routers/analytics.py: GET /api/analytics/reliability
- PatternEditor.tsx: ReliabilityBadge sur chaque card + usePatternReliability hook
- useApi.ts: usePatternReliability, useCalibration hooks
Sprint 2.2 — Contre-thèses & Invalidation Triggers
- database.py: migration ALTER TABLE — counter_thesis, invalidation_trigger, invalidation_probability
- database.py: save_custom_pattern() persiste les 3 nouveaux champs
- ai_analyzer.py: counter_thesis + invalidation_trigger + invalidation_probability dans le JSON schema
- auto_cycle.py: détection automatique des triggers d'invalidation contre les news (keyword match)
- routers/analytics.py: GET /api/analytics/invalidation-alerts
- PatternEditor.tsx: affichage contre-thèse dans les cards + champs dans le formulaire
- PatternEditor.tsx: affichage dans AiSuggestModal (suggestions IA)
- routers/patterns.py: PatternRequest inclut les 3 nouveaux champs
Sprint 2.3 — Calibration probabiliste & Demi-vie KB
- database.py: migration — predicted_probability sur pattern_score_history
- database.py: save_pattern_scores() stocke probability du pattern à chaque scoring run
- database.py: get_calibration_data() — Brier score + buckets de calibration par décile
- database.py: expires_at + confidence_decay_days sur knowledge_base
- database.py: decay_kb_confidence() — decay automatique + archivage à 0
- auto_cycle.py: decay_kb_confidence() appelé au début de chaque cycle (non-bloquant)
- routers/analytics.py: GET /api/analytics/calibration + POST /api/analytics/kb/decay
- frontend/src/pages/Analytics.tsx: nouvelle page — tableau fiabilité + calibration Brier
- App.tsx + Sidebar.tsx: route /analytics + entrée menu
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- ai_analyzer: add explicit calibration rules to SYSTEM_SCORER and batch
prompt so GPT-4o produces a spread of scores rather than defaulting
to 50 for all patterns (0-news patterns capped at 35, contra patterns
at 40, high-signal patterns can reach 70-85)
- auto_cycle: add _auto_synthesize_knowledge() called after each auto
portfolio snapshot; skips if last synthesis < 6h old to avoid
redundant GPT-4o calls — Super Contexte now updates automatically
every cycle without manual intervention
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
