Preprint · 2026
HiT-Forecast
Series-Level Routing over Foundation Time-Series Models with Hierarchical Transformer Meta-Features
1Bilkent University · 2Georgia Institute of Technology
Abstract
Time-series foundation models are now competitive zero-shot forecasters, yet no single model dominates across data sources and horizons. Picking the right foundation model for a given target series is therefore an open problem. We pose this problem at the level of an individual series rather than at the level of an individual forecasting window. For each series, the first 80% of its windows are run once through a pool of frozen foundation encoders to obtain pooled per-window embeddings; a meta-learner reads these embeddings, possibly together with the same-prefix history-error vector of the experts, and outputs a single routing distribution that is then applied to every future window of the same series. We implement this framework as HiT-Forecast, a hierarchical transformer encoder-pooling router whose tokens are per-window embeddings. The same architecture covers a linear-probe and tree-based meta-learner as constrained baselines, allowing a controlled feature ablation: the embeddings-plus-history arm recovers 50.0% of the best-static-to-oracle MASE gap on M4-hourly, compared with 29.5% for the feature-free history-argmin baseline and 8.9% for a linear probe on embeddings alone, with similar margins on M4-daily and M4-weekly. Complementing this positive result, a feature-free signal-ceiling analysis on three standard long-term-forecasting caches shows that the per-window argmin target on these benchmarks is statistically indistinguishable from min-of-K noise, which bounds what any per-window router on these caches can deliver and motivates the series-level reframing.
Method
A model-specific affine map projects each pooled embedding to a shared router dimension of 256, and a learnable identity embedding marks the expert. One trained router serves hard selection, soft weighted ensembling and a continuous combiner; the tables report hard selection. Routing quality is the share of the best-static-to-oracle MASE gap recovered, so best-static is 0% and the per-series oracle 100%.
Results
Recovers 50.0% of the oracle gap on M4-hourly
| Method | M4-hourly | M4-daily | M4-weekly |
|---|---|---|---|
| Oracle (per series) | 100.0 | 100.0 | 100.0 |
| Best-static expert | 0.0 | 0.0 | 0.0 |
| History-argmin | 29.5 | 28.6 | 18.2 |
| MLP (attention pooling) | 32.8 | 32.5 | 22.1 |
| HiT-Forecast (emb only) | 37.0 | 37.0 | 26.0 |
| HiT-Forecast (emb + history) | 50.0 | 45.5 | 32.5 |
Embeddings and error history are complementary for every meta-learner
| Meta-learner | History | Embeddings | Both |
|---|---|---|---|
| Linear probe | 23.9 | 8.9 | 32.8 |
| LightGBM | 26.4 | 11.5 | 38.7 |
| MLP (mean pooling) | 29.5 | 19.6 | 43.6 |
| HiT-Forecast | 29.5 | 32.8 | 50.0 |
The per-window target is indistinguishable from min-of-K noise
| Cache | Observed gap | Null gap | Persistence | Chance | Trailing |
|---|---|---|---|---|---|
| M4-hourly, per window | 0.063 | 0.142 | 0.412 | 0.418 | -2.7 |
| ETTh1, per window | 0.085 | 0.181 | 0.168 | 0.167 | -1.0 |
| Traffic, per window | 0.094 | 0.207 | 0.503 | 0.498 | -3.4 |
| M4-hourly, per series | 0.046 | 0.038 | 0.682 | 0.487 | 29.5 |
BibTeX
@unpublished{karaca2026hitforecast,
author = {Huseyin Karaca and A. Samil Namli and Suleyman Serdar Kozat},
title = {HiT-Forecast: Series-Level Routing over Foundation
Time-Series Models with Hierarchical Transformer
Meta-Features},
year = {2026},
note = {Preprint},
url = {https://hkaraca.com/hit-forecast/}
}