Current Data Inventory

Everything this model consumes lives in Supabase project sstmupvotyzrjygqoany (ap-northeast-1) and is served through the MCP gateway (Railway, 84 tools, 19 datasets).

Core tables

Table	Rows	Coverage	Refresh
ohlcv_1d	~23 908	81 symbols	Daily 06:00 UTC
ohlcv_1h	—	40 symbols	Hourly 0 0-21 * * *
ohlcv_15m	—	40 symbols	Hourly
ohlcv_5m	—	40 symbols	Hourly
tickers	116	15-min quotes	*/15 * * * *
economic_calendar	18	High-impact events	Daily
economic_indicators	~25 768	85 FRED series	Daily
central_bank_rates	—	G10 rates	Daily
cot_reports	~1 420	Commitment of Traders	Weekly
options_chains	~2 492	Select equities + crypto	Daily
equity_fundamentals	—	S&P 500 subset	Quarterly
earnings_estimates	~360	S&P 500 subset	Quarterly
news_articles	~13 855	6 RSS feeds · VADER	*/30 * * * *
ml_predictions	—	Kronos outputs	Every cron tick
signal_evaluations	—	Prediction labels	Hourly / daily

Gateway datasets

Per project memory, 19 registered datasets. The model is directly powered by:

• ohlcv_daily, ohlcv_intraday — candle source for Kronos input
• economic_calendar — event flags
• economic_indicators — macro context
• predictions — cached Kronos outputs served to UI

What the model consumes today

flowchart LR
    subgraph Existing
        OHLCV[(ohlcv_1h<br/>ohlcv_1d)] --> BATCH[kronos-batch-predict.py]
        CAL[(economic_calendar<br/>10-ch placeholder)] --> BATCH
    end
    BATCH --> ML[(ml_predictions)]
    ML --> GW[Gateway<br/>/showcase/ml-prediction]
    GW --> UI[prediction.datfxlabs.com]

What the model will need (Phase 1 · 20 channels)

See Data Gaps and Backfill Plan — both populated from the data-gap researcher report.

The short version:

• Have: economic_calendar with actual / forecast fields for primary events.
• Partial: surprise z-scores (need to compute rolling-20 std per event type at query time or materialise a view).
• Missing: FOMC hawkish score (requires NLP over statement text not currently stored), cross-asset leader OHLCV for DXY + VIX at 1 h resolution (verify coverage), FOMC statement text archive.

Data-quality guarantees

• No fake data — enforced by CLAUDE.md Data Integrity rules plus DB triggers blocking sources like 'fake', 'mock', 'simulated'.
• Source whitelist — only Yahoo Finance, FRED, Alpha Vantage, CoinGecko, Binance, Bybit, Finnhub, FMP, Marketaux.
• External verification — cross-checks against live market prices on data entry.