Preprint · 2026

HiT-Forecast

Series-Level Routing over Foundation Time-Series Models with Hierarchical Transformer Meta-Features

Huseyin Karaca1, A. Samil Namli2, Suleyman S. Kozat1

1Bilkent University · 2Georgia Institute of Technology

A series is split into eight history windows and two future windows. Every history window runs once through three frozen foundation encoders, Chronos-Large, Moirai-Base and Lag-Llama; each encoder's patch states for a window are mean-pooled to one token, giving one token per history window for each expert. Inside the HiT-Forecast router, a temporal encoder per expert summarises each token sequence, and a cross-expert fusion step, which also receives the experts' history errors, turns the summaries into routing weights. One argmax per series picks Moirai, which forecasts both future windows; Chronos and Lag-Llama are not run there.
The router reads one pooled encoder token per history window from every expert, plus their history errors; the one expert it picks forecasts every future window of the series.

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

MethodM4-hourlyM4-dailyM4-weekly
Oracle (per series)100.0100.0100.0
Best-static expert0.00.00.0
History-argmin29.528.618.2
MLP (attention pooling)32.832.522.1
HiT-Forecast (emb only)37.037.026.0
HiT-Forecast (emb + history)50.045.532.5
Table II. Oracle gap recovered (%), mean over five seeds that each re-derive the split over series.

Embeddings and error history are complementary for every meta-learner

Meta-learnerHistoryEmbeddingsBoth
Linear probe23.98.932.8
LightGBM26.411.538.7
MLP (mean pooling)29.519.643.6
HiT-Forecast29.532.850.0
Table III. Oracle gap recovered (%) on M4-hourly. Embeddings alone are weaker than the error history for three of the four families; together they add 8.9 to 20.5 points over history alone.

The per-window target is indistinguishable from min-of-K noise

CacheObserved gapNull gapPersistenceChanceTrailing
M4-hourly, per window0.0630.1420.4120.418-2.7
ETTh1, per window0.0850.1810.1680.167-1.0
Traffic, per window0.0940.2070.5030.498-3.4
M4-hourly, per series0.0460.0380.6820.48729.5
Table V. Permuting each expert's error column within a series reproduces the per-window gap (p = 1.00), and winner persistence sits at chance. On the series target the gap beats the null (p < 0.01) and a trailing-error router already recovers 29.5%. This bounds per-window routing on these caches, not in general.

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/}
}