Methodology and validation
Data sources & timing
Prices are Yahoo Finance daily bars (not intraday). Financials/filings come from Korea's DART and the US EDGAR. News comes from several RSS sources per market. The table below is read from the actual batch config (deploy/crontab.txt).
| Market | News sources | Post-close batch time (UTC / KST) |
|---|---|---|
| US | 8 | 22:30 / 07:30 |
| Korea | 6 | 07:15 / 16:15 |
| Crypto | 4 | 00:15 / 09:15 |
| commodity | 0 | 23:20 / 08:20 |
Point-in-time — no future leakage
Financial verdicts use only the quarter that was actually filed by that date — Korea by DART receipt date (rcept_no), US by EDGAR filed date. If a quarter was filed more than once, only the first filing is used; later restatements are excluded because they weren't known at the time (engine/quality_backtest.py). Price signals likewise use only closes available up to that day.
Signals and scoring
The sharp-drop risk score turns 17 signals (breadth, volatility regime, trend break, volume spikes, etc.) into percentiles (0-1, expanding window) using only data available up to that day, then combines them with a model retrained every January on prior data only, giving a 0-100 score (engine/crash.py). Stock scores are a walk-forward cross-sectional rank of trend/momentum/etc. factors.
Validation method
Walk-forward (no future data in training) + block-bootstrap 95% confidence intervals (holding periods overlap, so days aren't independent) + no conclusions drawn below n=30. The sharp-drop risk score's "Danger" stage is only used for alerts once a market has at least 20 sample days and a post-danger drawdown lift of at least 1.5x; markets that don't clear this bar still show the score, but alerts cap out at "Alert".
Limits — nothing hidden
- Forecast record length by market (actual days): US 40 · Korea 41 · Crypto 44 · commodity 0
- Daily bars only — intraday moves aren't captured.
- Execution, commissions and taxes aren't modeled (backtest cost assumptions are simulation estimates).
- Survivorship-bias handling is partial — we extended price history from 4 to 10 years (2026-09-20), but still reuse today's stock list across the whole past period. Only about 82%% of today's Korean names and 80%% of the US names already existed 10 years ago, so companies delisted along the way are missing. We have no point-in-time membership records, so we cannot separate how much of any increase comes from the longer sample and how much from looking only at survivors — we say so rather than guess.
- Sector returns are calculated by applying market-cap weights fixed at a single date to past periods unchanged — they don't match the weights that actually applied back then.
- Markets/stages with a short record withhold conclusions (see table above).
- Not investment advice — statistical information only.
Update frequency
Refreshed automatically 1-2 times a day after market close. This page last checked: 2026-11-19.