Method
How the forecasts on this site are made and scored. The figures below come from the site's own configuration when it is built.
Data
Daily candles (open, high, low, close, volume) in UTC, from each exchange's public market-data
endpoint through the open-source ccxt library: no account, no API key, no trading endpoint. Each
asset comes from one exchange (Binance, KuCoin, MEXC or Kraken), priced in USDT or USD. Only days
that have closed are stored, and stored days are never changed.
Every refresh validates each series for gaps, duplicate days, stale prices, impossible candles (a low above a high, for instance) and jumps where one trading symbol replaced another. A finding that is understood, such as an exchange pausing a market, is documented with its cause; an unexplained one stops that asset's forecast for the day.
Models
- TimesFM: Google's pretrained time-series foundation model, here timesfm-3.0, with its checkpoint pinned to one revision. It sees only the asset's past closes.
- Naive: every future close equals the last one. The baseline a forecast must beat.
- Seasonal naive: each future day repeats the close of the same weekday a week before.
- AutoETS and AutoARIMA: classical statistical models, fitted to each input with a
weekly season, from the open-source
statsforecastlibrary.
Each model gives nine deciles (10% to 90%) for each of the next 30 days. The median is the point forecast; the 30–70% and 10–90% ranges are the 40% and 80% intervals.
Walk-forward, no lookahead
A forecast made on day c (its cutoff) uses only closes dated before c: at most 1024 days of them, and none if fewer than 128 remain. Gaps of up to 7 days repeat the last close; after a longer gap the input starts again. The same rule applies to every model and every asset.
Live forecasts are made once a day, after the previous day's close (UTC), and stored the moment they are made. A stored forecast is never changed, and a day's forecast is only valid if made before that day ends: its first target is that day's close.
Scoring
A forecast is scored once its target day has closed. For each model, asset and horizon:
- MASE: the mean absolute error, scaled by the input's mean absolute day-to-day change, so assets of any price compare.
- Relative error: the absolute error as a share of the last input close.
- Direction: the share of up-or-down calls that went the right way. A forecast of no move, or a day with no move, isn't counted, so naive has no score.
- Coverage: the share of closes inside each interval; a calibrated 80% interval holds 80%.
Assets are grouped in tiers (control: stablecoins and EUR; large, mid and micro caps). A tier's figure is the mean of its assets' figures, each asset weighted equally.
Verdicts
Whether TimesFM beats naive is judged per tier at horizons 1, 7, 30, from the ratio of their MASE. Forecasts from neighbouring days share most of their input, so they aren't independent: a moving-block bootstrap resamples whole 30-day blocks of forecast days, the same days for every asset in the tier, 1000 times. If the 90% interval of the ratio lies below 1, TimesFM is better; above 1, worse; otherwise there is no detectable difference.
The live record gives a verdict only once a tier has 180 scored forecast days at that horizon, a number fixed before the first live result. Until then it says "too early to tell".
Missed days
The daily job runs on one PC. A day it doesn't run, or runs too late, is a missed day: it is counted and listed on the record page, and never filled in afterwards, because a forecast made later could see data its day couldn't have.
The backtest
Before going live, the same models and rules were replayed over the past: a forecast from every day with enough history, each from the data before it only. That replay is the backtest record. It shows how the models would have done; the live record shows how they do.