# Alphaforecast

Quantile price forecasts for Bittensor subnet alpha tokens and TAO.
Model: tensorlink-dev/yumoto-alpha-v0.1-22m. Republished every ~15 minutes.

## Fetching

All endpoints are public, CORS-open (access-control-allow-origin: *), no key.

- /data/latest.json      every subnet, close-price deciles at 1h/6h/24h/168h
- /data/tokens.json      index: subnets, timeframes, per-timeframe accuracy
- /data/sn{netuid}_{tf}.csv  full series, tf in 5m|1h|4h|1d

Start with latest.json. Fetch a CSV only if you need the full horizon,
context history, or the open/high/low quantiles.

## CSV schema

  role      "context" (observed) or "forecast" (model output)
  step      bar index; forecast rows start at 0
  ts        unix seconds, bar open
  o/h/l/c   observed OHLC, context rows only
  {target}_q{level}  quantile, target in close|open|high|low,
            level in 10..90 (deciles). q50 is the median.

Forecast rows leave the observed OHLC columns empty, and vice versa.

## Caveats

Forecasts are model output, not advice, and nobody stands behind them.
Subnet alpha tokens are thinly traded and can move violently or delist.
accuracy.geomean_ratio in tokens.json is the ratio to a naive baseline --
below 1 means the model beat it, and it is close to 1 on some timeframes.
Past scores do not carry forward.