Datasets:
Add tabSutra_39M_v1.0_rules (history + structured effects, deterministic rules)
tabSutra_39M_v1.0_rules
This pull request adds the folder eval/tabSutra_39M_v1.0_rules/ with tabSutra_39M_v1.0_rules.parquet and config.json.
The method is our history-only model, tabSutra_39M_v1.0, which is already submitted for the leaderboard, with its forecasts edited
by three deterministic rules that read the structured effects. There is no LLM and no agent in the pipeline, and the
model itself is unchanged. The category is "TSFM + rules" and the input is "History + Macro + Micro".
How the rule set was chosen
We split the 10,000 instances into two fixed halves, stratified by dataset and term, so that both halves contain every
group. We scored twelve candidate rule sets on half A and chose the best one. We then checked that the same rule set was
also the best on half B. It was: clip + mean_shift + excursion scores 0.639 on half A, 0.614 on half B, and 0.627 on
all instances. The candidates that invent a magnitude the schema does not give, namely a ramp for the trend, a rescale
of the seasonal component through a moving-average split, and an affine match of the path to level_range, scored
worse on both halves and were dropped.
For comparison, the same three rules applied to a seasonal naive path score 0.768, and applied to the last observed
value they score 0.870.
The rules
Each rule applies the benchmark's own definition of an effect, using the benchmark's level normalisation, in which a
level ℓ stands for the raw value history_min + ℓ / 100 × (history_max − history_min).
- The path is clipped to the interval given by level_range.
- For a mean_shift, the path ramps between start and end to the settled level given by intensity, and it stays
at that level for the rest of the horizon. - For an excursion of subtype impulse or sustained, a bump is added between start and end. The bump is zero at
both ends and is sized so that the maximum of the interval (for direction up) or its minimum (for direction down)
equals the level given by intensity. If the path already goes beyond that level, it is clipped to it.
The remaining effects are not used: the trend segments, the macro seasonality, every seasonal_change, recurrent
excursions, and the fields amplitude_trend and envelope_change. The reason is that only the three effects above
fix a numerical target in the schema. A trend segment gives a direction and a shape but no size, a recurrent excursion
gives a level but no event times, and a seasonal change needs a separation of level and season that the schema does not
define.
Inputs read
The pipeline reads history, n_horizon, frequency, macro and micro. It does not read future or
tsfm_average_forecast. The timestamps given to the model are the same synthetic ones as in our history-only
submission: row k is placed at k × step from an origin of zero at the benchmark's frequency.
MASE
The mase column was computed with the benchmark's own metrics.py code.
Contact: https://tabsutra.com