enigma.lt

Record

Two records, never mixed: the live record scores the forecasts this site published, as their target days close; the backtest record replays the past with the same models and rules. Figures are per tier: the mean over its assets, each weighted equally.

Live record

1 daily forecast since 2026-11-15, scored as of 2026-11-15. No day missed.

Does TimesFM 3.0 beat naive?

The ratio of their MASE per tier and horizon, with a 90% interval from a block bootstrap over forecast days. Below 1 means TimesFM 3.0's errors are smaller. A verdict needs 180 scored forecast days, a number fixed before the first live result.

TierHorizonScored days Ratio90% intervalVerdict
control10 of 180 – – too early to tell
control70 of 180 – – too early to tell
control300 of 180 – – too early to tell
large10 of 180 – – too early to tell
large70 of 180 – – too early to tell
large300 of 180 – – too early to tell
mid10 of 180 – – too early to tell
mid70 of 180 – – too early to tell
mid300 of 180 – – too early to tell
micro10 of 180 – – too early to tell
micro70 of 180 – – too early to tell
micro300 of 180 – – too early to tell

Every live forecast, as stored: live-store.csv.

Backtest record

A replay, not live forecasts: every day of the past, each model forecast from the data before that day only. Run backtest_2026-10-21_20261026T160033Z, data to 2026-10-21, enigma 6c03bbb, timesfm 3.0.2, torch 2.14.1+cpu, statsforecast 2.1.1.

Verdicts

TierHorizonModels Ratio90% intervalVerdict Forecasts
control1timesfm-2.5 / naive 0.9990.986 – 1.010 no detectable difference5308
control1timesfm-3.0 / naive 0.9820.972 – 0.990 better5308
control1timesfm-3.0 / timesfm-2.5 0.9830.977 – 0.989 better5308
control7timesfm-2.5 / naive 0.9820.961 – 1.008 no detectable difference5302
control7timesfm-3.0 / naive 0.9800.957 – 1.007 no detectable difference5302
control7timesfm-3.0 / timesfm-2.5 0.9980.986 – 1.012 no detectable difference5302
control30timesfm-2.5 / naive 0.9760.929 – 1.037 no detectable difference5279
control30timesfm-3.0 / naive 1.0070.958 – 1.064 no detectable difference5279
control30timesfm-3.0 / timesfm-2.5 1.0310.995 – 1.066 no detectable difference5279
large1timesfm-2.5 / naive 1.0351.024 – 1.046 worse15947
large1timesfm-3.0 / naive 1.0161.012 – 1.019 worse15947
large1timesfm-3.0 / timesfm-2.5 0.9820.972 – 0.991 better15947
large7timesfm-2.5 / naive 1.0601.028 – 1.085 worse15947
large7timesfm-3.0 / naive 1.0311.019 – 1.042 worse15947
large7timesfm-3.0 / timesfm-2.5 0.9730.952 – 0.994 better15947
large30timesfm-2.5 / naive 1.0441.006 – 1.077 worse15947
large30timesfm-3.0 / naive 1.0341.013 – 1.052 worse15947
large30timesfm-3.0 / timesfm-2.5 0.9910.972 – 1.013 no detectable difference15947
mid1timesfm-2.5 / naive 1.0241.013 – 1.034 worse4742
mid1timesfm-3.0 / naive 1.0121.004 – 1.020 worse4742
mid1timesfm-3.0 / timesfm-2.5 0.9890.979 – 0.997 better4742
mid7timesfm-2.5 / naive 1.0601.036 – 1.088 worse4742
mid7timesfm-3.0 / naive 1.0291.006 – 1.049 worse4742
mid7timesfm-3.0 / timesfm-2.5 0.9710.936 – 1.000 better4742
mid30timesfm-2.5 / naive 1.1091.047 – 1.178 worse4742
mid30timesfm-3.0 / naive 1.0550.996 – 1.119 no detectable difference4742
mid30timesfm-3.0 / timesfm-2.5 0.9520.869 – 1.019 no detectable difference4742
micro1timesfm-2.5 / naive 1.0161.008 – 1.026 worse6453
micro1timesfm-3.0 / naive 1.0070.998 – 1.016 no detectable difference6453
micro1timesfm-3.0 / timesfm-2.5 0.9910.982 – 0.999 better6453
micro7timesfm-2.5 / naive 1.0231.006 – 1.044 worse6453
micro7timesfm-3.0 / naive 1.0070.993 – 1.020 no detectable difference6453
micro7timesfm-3.0 / timesfm-2.5 0.9850.971 – 0.997 better6453
micro30timesfm-2.5 / naive 1.0330.992 – 1.080 no detectable difference6453
micro30timesfm-3.0 / naive 1.0050.967 – 1.046 no detectable difference6453
micro30timesfm-3.0 / timesfm-2.5 0.9730.930 – 1.011 no detectable difference6453

Every model, backtest

MASE (lower is better; the lowest per row in bold)

TierHorizon timesfm-3.0timesfm-2.5auto-arimaauto-etsnaiveseasonal-naive
control1 0.703 0.715 0.754 0.715 0.716 1.577
control7 1.545 1.547 1.841 1.577 1.576 1.576
control30 2.748 2.664 4.062 2.727 2.730 2.916
large1 1.443 1.469 1.468 1.423 1.420 3.871
large7 4.089 4.202 4.114 3.998 3.965 3.965
large30 9.853 9.947 10.369 9.842 9.527 10.409
mid1 0.764 0.772 0.767 0.757 0.755 2.058
mid7 2.130 2.195 2.152 2.071 2.070 2.070
mid30 4.787 5.029 4.842 4.543 4.537 5.040
micro1 0.790 0.798 0.822 0.795 0.785 2.234
micro7 2.252 2.287 2.393 2.252 2.237 2.237
micro30 4.705 4.835 5.379 4.740 4.680 5.160

Relative error (absolute error as a share of the last close; lowest in bold)

TierHorizon timesfm-3.0timesfm-2.5auto-arimaauto-etsnaiveseasonal-naive
control1 0% 0% 0% 0% 0% 0%
control7 0% 0% 0% 0% 0% 0%
control30 1% 1% 1% 1% 1% 1%
large1 3% 3% 3% 3% 3% 9%
large7 9% 9% 9% 9% 9% 9%
large30 23% 23% 23% 23% 23% 25%
mid1 4% 4% 4% 3% 3% 10%
mid7 10% 10% 10% 10% 10% 10%
mid30 23% 23% 24% 21% 21% 23%
micro1 4% 4% 4% 4% 4% 11%
micro7 11% 12% 12% 11% 11% 11%
micro30 25% 26% 29% 25% 25% 27%

Direction (share of up or down calls that were right; naive makes none)

TierHorizon timesfm-3.0timesfm-2.5auto-arimaauto-etsnaiveseasonal-naive
control1 57% 56% 54% 55% – 56%
control7 57% 57% 53% 52% – –
control30 61% 62% 56% 53% – 57%
large1 49% 49% 50% 52% – 51%
large7 49% 50% 49% 51% – –
large30 50% 52% 50% 50% – 48%
mid1 51% 51% 46% 51% – 50%
mid7 50% 48% 46% 51% – –
mid30 50% 45% 45% 50% – 49%
micro1 51% 54% 52% 51% – 50%
micro7 51% 50% 52% 50% – –
micro30 50% 48% 51% 51% – 49%

80% interval coverage (calibrated at 80%)

TierHorizon timesfm-3.0timesfm-2.5auto-arimaauto-etsnaiveseasonal-naive
control1 80% 87% 92% 93% 93% 91%
control7 80% 81% 90% 94% 96% 91%
control30 82% 83% 87% 96% 97% 94%
large1 81% 87% 83% 87% 84% 82%
large7 81% 80% 81% 84% 83% 82%
large30 78% 78% 78% 80% 80% 79%
mid1 84% 90% 91% 86% 91% 90%
mid7 80% 82% 90% 85% 91% 90%
mid30 75% 77% 89% 84% 89% 88%
micro1 81% 87% 93% 89% 93% 93%
micro7 79% 79% 92% 86% 92% 93%
micro30 78% 78% 92% 86% 92% 92%