BTC/USDTHow the forecast is made
- The model. TimesFM 2.5 with 200 million parameters (google/timesfm-2.5-200m-pytorch), run on CPU. It reads up to 1,024 closed candles of a coin, the log of the price, and returns 10 quantiles for every future candle. No news, no order book, no other coins: price history only.
- The coins. The 30 most traded USDT pairs on Binance for the 7-day view and the 12 most traded for the 24-hour view, stablecoins, wrapped and leveraged tokens left out. Bitcoin is always included.
- Chance higher. Where today's price falls inside the forecast's quantiles at the horizon. 60% or more counts as "leans up", 40% or less as "leans down". It never shows 0% or 100%, because the model gives no view beyond its 10-90% range.
- Every 4 hours on the publisher's computer, after each 4-hour candle closes. The daily view changes once a day, at 00:00 UTC.
How good is it? The walk-forward test
At every test date each coin got the forecast the live page would have shown that day, from candles closed before it, and the result was checked against what the price did next. Direction is compared with a coin toss and with always saying "up"; size with "no change"; the 10-90% range with its promised 80%. The rank test buys the 5 coins with the best forecast, sells at the horizon and pays 0.1% a side. Today's most traded coins survived to be on the list, which flatters every number that just holds coins (survivorship); the rank test compares coins of the same list on the same date, which cancels most of it.