Write-up
Can a time-series foundation model forecast daily crypto prices better than "tomorrow's close equals today's"? This page answers the project's four questions from two backtests: replays of 2017–2026 in which every model forecast each day from the prices before it only. They are a replay, not a live record, and their findings are provisional. The tables below are generated from the backtests' own result files whenever this site is built.
- Backtest 1 (run
backtest_2026-10-21_20261026T160033Z, data to 2026-10-21): TimesFM 2.5 and 3.0 against four classical baselines on 17 assets. - Backtest 2 (run
backtest-cross-asset_2026-10-21_20261104T072655Z, data to 2026-10-21): TimesFM 3.0 also given the past closes of Bitcoin and Ether.
In short
- Q1 · Does TimesFM beat naive? No. On large and mid caps both versions are slightly but detectably worse than naive at almost every horizon. The one win is TimesFM 3.0 on stablecoins at one day ahead, by 1.8%.
- Q2 · Are its quantile forecasts calibrated? Yes, far better than the baselines'. Its intervals are about as wide as they should be; the baselines' are mostly much too wide.
- Q3 · How does this change across tiers? The loss to naive is clearest on large and mid caps. On micro caps and stablecoins TimesFM is mostly indistinguishable from naive.
- Q4 · Do other coins' histories help? No detectable gain from Bitcoin's and Ether's past. Coins move with Bitcoin on the same day, not after it, and a forecast can't see the same day.
So for prices, TimesFM's point forecast is no better than a random walk, which is what you would expect if crypto prices are close to one. Its sense of uncertainty, though, is well judged.
Q1 · Does TimesFM beat naive?
Naive forecasts that every future close equals the last one. The table gives the ratio of each TimesFM version's error (MASE) to naive's, per tier and horizon (h, in days), with a 90% interval from a bootstrap over forecast days. Below 1 means TimesFM's errors are smaller; "better" or "worse" is claimed only when the whole interval is on one side of 1.
| tier | h | ratio of | ratio | interval | verdict | assets | pairs |
|---|---|---|---|---|---|---|---|
| control | 1 | timesfm-2.5 / naive | 0.999 | [0.986, 1.010] | no detectable difference | 3 | 5308 |
| control | 1 | timesfm-3.0 / naive | 0.982 | [0.972, 0.990] | better | 3 | 5308 |
| control | 7 | timesfm-2.5 / naive | 0.982 | [0.961, 1.008] | no detectable difference | 3 | 5302 |
| control | 7 | timesfm-3.0 / naive | 0.980 | [0.957, 1.007] | no detectable difference | 3 | 5302 |
| control | 30 | timesfm-2.5 / naive | 0.976 | [0.929, 1.037] | no detectable difference | 3 | 5279 |
| control | 30 | timesfm-3.0 / naive | 1.007 | [0.958, 1.064] | no detectable difference | 3 | 5279 |
| large | 1 | timesfm-2.5 / naive | 1.035 | [1.024, 1.046] | worse | 6 | 15947 |
| large | 1 | timesfm-3.0 / naive | 1.016 | [1.012, 1.019] | worse | 6 | 15947 |
| large | 7 | timesfm-2.5 / naive | 1.060 | [1.028, 1.085] | worse | 6 | 15947 |
| large | 7 | timesfm-3.0 / naive | 1.031 | [1.019, 1.042] | worse | 6 | 15947 |
| large | 30 | timesfm-2.5 / naive | 1.044 | [1.006, 1.077] | worse | 6 | 15947 |
| large | 30 | timesfm-3.0 / naive | 1.034 | [1.013, 1.052] | worse | 6 | 15947 |
| mid | 1 | timesfm-2.5 / naive | 1.024 | [1.013, 1.034] | worse | 4 | 4742 |
| mid | 1 | timesfm-3.0 / naive | 1.012 | [1.004, 1.020] | worse | 4 | 4742 |
| mid | 7 | timesfm-2.5 / naive | 1.060 | [1.036, 1.088] | worse | 4 | 4742 |
| mid | 7 | timesfm-3.0 / naive | 1.029 | [1.006, 1.049] | worse | 4 | 4742 |
| mid | 30 | timesfm-2.5 / naive | 1.109 | [1.047, 1.178] | worse | 4 | 4742 |
| mid | 30 | timesfm-3.0 / naive | 1.055 | [0.996, 1.119] | no detectable difference | 4 | 4742 |
| micro | 1 | timesfm-2.5 / naive | 1.016 | [1.008, 1.026] | worse | 4 | 6453 |
| micro | 1 | timesfm-3.0 / naive | 1.007 | [0.998, 1.016] | no detectable difference | 4 | 6453 |
| micro | 7 | timesfm-2.5 / naive | 1.023 | [1.006, 1.044] | worse | 4 | 6453 |
| micro | 7 | timesfm-3.0 / naive | 1.007 | [0.993, 1.020] | no detectable difference | 4 | 6453 |
| micro | 30 | timesfm-2.5 / naive | 1.033 | [0.992, 1.080] | no detectable difference | 4 | 6453 |
| micro | 30 | timesfm-3.0 / naive | 1.005 | [0.967, 1.046] | no detectable difference | 4 | 6453 |
On large and mid caps, the two versions are worse than naive in 11 of 12 cases, with errors 1–11% larger. No baseline beats naive either; AutoETS comes closest. Every model's error, the lowest per row in bold:
| tier | h | timesfm-2.5 | timesfm-3.0 | naive | seasonal-naive | auto-ets | auto-arima |
|---|---|---|---|---|---|---|---|
| control | 1 | 0.715 | 0.703 | 0.716 | 1.577 | 0.715 | 0.754 |
| control | 7 | 1.547 | 1.545 | 1.576 | 1.576 | 1.577 | 1.841 |
| control | 30 | 2.664 | 2.748 | 2.730 | 2.916 | 2.727 | 4.062 |
| large | 1 | 1.469 | 1.443 | 1.420 | 3.871 | 1.423 | 1.468 |
| large | 7 | 4.202 | 4.089 | 3.965 | 3.965 | 3.998 | 4.114 |
| large | 30 | 9.947 | 9.853 | 9.527 | 10.409 | 9.842 | 10.369 |
| mid | 1 | 0.772 | 0.764 | 0.755 | 2.058 | 0.757 | 0.767 |
| mid | 7 | 2.195 | 2.130 | 2.070 | 2.070 | 2.071 | 2.152 |
| mid | 30 | 5.029 | 4.787 | 4.537 | 5.040 | 4.543 | 4.842 |
| micro | 1 | 0.798 | 0.790 | 0.785 | 2.234 | 0.795 | 0.822 |
| micro | 7 | 2.287 | 2.252 | 2.237 | 2.237 | 2.252 | 2.393 |
| micro | 30 | 4.835 | 4.705 | 4.680 | 5.160 | 4.740 | 5.379 |
TimesFM 3.0 is detectably better than 2.5 in 7 of 12 tier and horizon combinations, and never detectably worse:
| tier | h | ratio of | ratio | interval | verdict | assets | pairs |
|---|---|---|---|---|---|---|---|
| control | 1 | timesfm-3.0 / timesfm-2.5 | 0.983 | [0.977, 0.989] | better | 3 | 5308 |
| control | 7 | timesfm-3.0 / timesfm-2.5 | 0.998 | [0.986, 1.012] | no detectable difference | 3 | 5302 |
| control | 30 | timesfm-3.0 / timesfm-2.5 | 1.031 | [0.995, 1.066] | no detectable difference | 3 | 5279 |
| large | 1 | timesfm-3.0 / timesfm-2.5 | 0.982 | [0.972, 0.991] | better | 6 | 15947 |
| large | 7 | timesfm-3.0 / timesfm-2.5 | 0.973 | [0.952, 0.994] | better | 6 | 15947 |
| large | 30 | timesfm-3.0 / timesfm-2.5 | 0.991 | [0.972, 1.013] | no detectable difference | 6 | 15947 |
| mid | 1 | timesfm-3.0 / timesfm-2.5 | 0.989 | [0.979, 0.997] | better | 4 | 4742 |
| mid | 7 | timesfm-3.0 / timesfm-2.5 | 0.971 | [0.936, 1.000] | better | 4 | 4742 |
| mid | 30 | timesfm-3.0 / timesfm-2.5 | 0.952 | [0.869, 1.019] | no detectable difference | 4 | 4742 |
| micro | 1 | timesfm-3.0 / timesfm-2.5 | 0.991 | [0.982, 0.999] | better | 4 | 6453 |
| micro | 7 | timesfm-3.0 / timesfm-2.5 | 0.985 | [0.971, 0.997] | better | 4 | 6453 |
| micro | 30 | timesfm-3.0 / timesfm-2.5 | 0.973 | [0.930, 1.011] | no detectable difference | 4 | 6453 |
Q2 · Are its quantile forecasts calibrated?
A calibrated 80% interval contains the actual close 80% of the time, and a 20% interval 20% of the time. The actual close lands inside TimesFM's 80% interval 75.5–90.5% of the time, and inside its 20% interval 18–31% of the time. The baselines' 20% intervals catch 23–59% of closes: they are too wide. TimesFM 3.0 is the closest to nominal.
| nominal | tier | h | timesfm-2.5 | timesfm-3.0 | naive | seasonal-naive | auto-ets | auto-arima |
|---|---|---|---|---|---|---|---|---|
| 20% | control | 1 | 25.8% | 23.1% | 48.0% | 41.3% | 45.5% | 43.5% |
| 20% | control | 7 | 19.8% | 20.3% | 52.3% | 41.5% | 44.3% | 39.3% |
| 20% | control | 30 | 23.9% | 20.2% | 58.6% | 50.0% | 52.1% | 36.2% |
| 20% | large | 1 | 26.0% | 20.7% | 40.9% | 38.8% | 31.9% | 40.4% |
| 20% | large | 7 | 20.9% | 22.5% | 39.7% | 38.8% | 27.6% | 38.3% |
| 20% | large | 30 | 24.3% | 20.8% | 35.3% | 34.1% | 22.8% | 33.6% |
| 20% | mid | 1 | 30.9% | 19.6% | 51.8% | 48.7% | 27.3% | 51.4% |
| 20% | mid | 7 | 21.1% | 21.0% | 49.7% | 48.9% | 25.0% | 49.8% |
| 20% | mid | 30 | 27.1% | 18.0% | 48.2% | 46.7% | 23.8% | 48.5% |
| 20% | micro | 1 | 29.5% | 19.3% | 57.1% | 50.6% | 35.9% | 53.8% |
| 20% | micro | 7 | 20.7% | 21.4% | 51.3% | 51.0% | 31.4% | 47.5% |
| 20% | micro | 30 | 25.1% | 21.0% | 49.1% | 47.8% | 28.8% | 46.5% |
| 80% | control | 1 | 87.2% | 80.4% | 93.0% | 91.2% | 92.8% | 92.5% |
| 80% | control | 7 | 80.6% | 80.4% | 95.7% | 91.1% | 93.9% | 90.5% |
| 80% | control | 30 | 83.2% | 81.9% | 96.8% | 94.1% | 95.6% | 87.3% |
| 80% | large | 1 | 87.1% | 80.7% | 83.5% | 82.1% | 87.2% | 82.6% |
| 80% | large | 7 | 80.4% | 80.6% | 82.7% | 81.9% | 84.1% | 81.3% |
| 80% | large | 30 | 77.6% | 78.1% | 79.9% | 78.7% | 80.4% | 78.3% |
| 80% | mid | 1 | 90.5% | 84.1% | 91.3% | 90.0% | 86.2% | 90.8% |
| 80% | mid | 7 | 82.0% | 80.0% | 90.7% | 90.1% | 85.1% | 89.8% |
| 80% | mid | 30 | 77.0% | 75.5% | 89.3% | 88.0% | 83.6% | 88.7% |
| 80% | micro | 1 | 87.4% | 80.8% | 93.0% | 92.9% | 88.7% | 92.7% |
| 80% | micro | 7 | 79.4% | 78.7% | 92.0% | 92.9% | 86.2% | 91.8% |
| 80% | micro | 30 | 77.6% | 77.8% | 92.0% | 92.1% | 86.0% | 91.8% |
Q3 · How does performance change across tiers?
The tiers run from stablecoin controls (USDC, USDT and EUR, which should barely move) through large and mid caps to micro caps. The loss to naive is clearest on large and mid caps (Q1's table). On micro caps and stablecoins TimesFM is mostly indistinguishable from naive. Stablecoins are where it does best, by predicting the drift back to the peg, which says little about markets that have no peg.
Q4 · Do other assets' histories help?
In the second backtest TimesFM 3.0 also got the closes of Bitcoin and Ether up to the day before each forecast. Its errors (here relative error, the error as a share of the last close) are the same as without them in 10 of 12 tier and horizon combinations, and slightly larger in the other 2: large caps at 1 and 7 days, by 0.2% and 0.8%. The intervals put any gain the two leaders might give below 0.5% at one day, below 1.7% at a week, and below 5% at 30 days.
| tier | h | ratio of | ratio | interval | verdict | assets | pairs |
|---|---|---|---|---|---|---|---|
| control | 1 | timesfm-3.0+lead / timesfm-3.0 | 1.000 | [0.995, 1.004] | no detectable difference | 3 | 5308 |
| control | 7 | timesfm-3.0+lead / timesfm-3.0 | 0.999 | [0.987, 1.012] | no detectable difference | 3 | 5302 |
| control | 30 | timesfm-3.0+lead / timesfm-3.0 | 0.993 | [0.964, 1.024] | no detectable difference | 3 | 5279 |
| large | 1 | timesfm-3.0+lead / timesfm-3.0 | 1.002 | [1.000, 1.004] | worse | 6 | 15947 |
| large | 7 | timesfm-3.0+lead / timesfm-3.0 | 1.008 | [1.001, 1.014] | worse | 6 | 15947 |
| large | 30 | timesfm-3.0+lead / timesfm-3.0 | 1.009 | [0.996, 1.024] | no detectable difference | 6 | 15947 |
| mid | 1 | timesfm-3.0+lead / timesfm-3.0 | 1.003 | [0.999, 1.008] | no detectable difference | 4 | 4742 |
| mid | 7 | timesfm-3.0+lead / timesfm-3.0 | 0.994 | [0.983, 1.006] | no detectable difference | 4 | 4742 |
| mid | 30 | timesfm-3.0+lead / timesfm-3.0 | 0.982 | [0.953, 1.008] | no detectable difference | 4 | 4742 |
| micro | 1 | timesfm-3.0+lead / timesfm-3.0 | 1.001 | [0.997, 1.004] | no detectable difference | 4 | 6453 |
| micro | 7 | timesfm-3.0+lead / timesfm-3.0 | 1.007 | [0.998, 1.015] | no detectable difference | 4 | 6453 |
| micro | 30 | timesfm-3.0+lead / timesfm-3.0 | 0.995 | [0.967, 1.021] | no detectable difference | 4 | 6453 |
Why no gain: a coin's daily return moves with Bitcoin's of the same day (a correlation of 0.41–0.65 by tier, outside stablecoins), but Bitcoin's return one to seven days earlier says almost nothing about it: at most 0.06 by tier. On the 3% of days that follow an unusually large Bitcoin move, where a lead should show most, no tier gains at any horizon, though those days are few. Five other coins as input, with or without the two leaders, did not help either. A forecast gain would have shown that one coin's moves lead another's, not that they cause them; there was none to show.
Caveats
- 36 verdicts at 90%. Even with no real differences, a few verdicts would come out "better" or "worse" by chance. The pattern, consistent losses on large and mid caps, matters more than any one row.
- MASE leans on trending periods. Its scale is the input's mean daily move over up to 1,024 days, so after a long rally one day's error can score 20–40. Every model shares the same scale for a forecast, so the ratios stay fair. The second backtest measured relative error instead, and the first one's finding held.
- Stablecoins are a weak control. USDC has no data from 2022-09-27 to 2023-03-10, USDT has only 720 days, from Kraken, and EUR/USDT is a crypto venue's view of the euro, not foreign exchange.
- One way of asking. Zero-shot, raw closes, a 1,024-day cap. Log prices, returns or longer inputs might change the picture; they were left out on purpose, so that no setting was tuned on these results.
- Overlapping evidence. With a forecast every day, 30-day forecasts overlap almost completely, so there are far fewer independent observations than forecasts. That is why the intervals are as wide as they are.
What the live record can and cannot show yet
Live forecasts started on 2026-11-15; 1 day so far. Each is stored the moment it is made and never changed, so it is a forecast of a future that had not happened yet, not a replay. That is what the live record adds to the backtests.
It cannot yet say whether TimesFM beats naive. A tier gets a verdict only after 180 scored forecast days at a horizon, a number fixed before the first live result. If no day is missed, the first possible verdicts come on 2027-05-14 at one day ahead and on 2027-06-12 at 30 days. Until then the record shows the counts and every model's figures, but a few weeks of a few assets are too little to tell models apart. Nor are the live figures comparable number for number with the backtests': they come from a different and much shorter period.