TimeTron-v2-33M

A 33.5M-parameter time-series foundation model, distilled from Chronos-2 by latent-space knowledge distillation — the student matches the teacher's internal representations rather than its forecasts.

GIFT-Eval: 0.8794 normalized MASE / 0.6122 normalized CRPS across all 97 configurations (official Salesforce harness). On probabilistic accuracy it beats models up to 21× larger.

Trained on 8.25B points (7.5B latent distillation + 750M output distillation) on a single rented RTX 4090.

Usage

import numpy as np
from modeling_timetron import TimeTron

model = TimeTron.from_pretrained("CorteriIntelligence/TimeTron-v2-33M")

context = np.random.randn(4, 512).astype("float32")     # (batch, history)
quantiles = model.predict(context, prediction_length=128)   # (4, 128, 21), input's own scale
median = model.predict_median(context, prediction_length=128)  # (4, 128)

• 21 quantile levels: [0.01, 0.05, 0.1, 0.15 … 0.9, 0.95, 0.99] — index 10 is the median
• 384 steps decoded natively; longer horizons extend autoregressively in whole chunks
• Any context length — cropped to a multiple of 32 and left-padded as needed, up to 2048
• Output is in the input's own scale; no manual normalization required

Requires torch, numpy, and safetensors. modeling_timetron.py is self-contained.

Model details

Forecasts are produced in z = asinh((y − μ)/σ) space and inverted as ŷ = μ + σ·sinh(z), which is why the model handles wide dynamic ranges without manual scaling.

Training

Latent KD at three taps (student layers 4/8/12 → teacher layers 4/8/final) with SmoothL1 against LayerNorm'd, pooled teacher representations, plus a pinball loss on 21 quantiles and a masked-view consistency term. The final checkpoint is a weight average of one latent-distilled checkpoint and two output-distilled variants (W_OUT 0.2 and 0.5) — averaging over objective-diverse checkpoints was the single most reliable source of gain in the project.

Limitations

• Test-data leakage: declared Yes. Training pools contain series that are GIFT-Eval test datasets (m4 family, LOOP_SEATTLE, SZ_TAXI, hierarchical_sales, restaurant, temperature_rain). Long series used a 10% tail holdout and short series a last-window holdout, but partial overlap with official test horizons remains possible. We declare it rather than argue the edge case.
• Univariate. The architecture contains a group-attention branch for in-context learning across related series, and predict(..., group_ids=...) exposes it — but on this checkpoint it is untrained and passing group_ids is a no-op. Training it on a frozen backbone measured 3.4% worse on GIFT's 43 multivariate configurations; the capability turned out to be inseparable from the backbone it co-adapts with.
• Known-future covariates are supported by the architecture (future_values) with zero additional parameters, but are only lightly trained (~5% of rows).
• Trained at context ≤512. Longer contexts are accepted but untested.
• Horizons beyond 384 use autoregressive chunk extension, not a native long-horizon head.

Evaluation

Numbers above are the official GIFT-Eval harness (gluonts metric engine, official windowing and seasonality), aggregated as the geometric mean of per-configuration model_MASE / seasonal_naive_MASE. Submission artifacts are in gifteval_submission/.

On IEX Indian electricity spot prices (private data, clean for every model compared), TimeTron scores 1.189 MASE / 862 CRPS versus TimesFM-2.5's 1.232 / 908 — better on both at 1/7 the size.

Citation

@misc{srivastava2026timetron,
  title  = {TimeTron: Latent-Space Distillation of a Time-Series Foundation Model at 33M Parameters},
  author = {Srivastava, Aditya},
  year   = {2026},
  note   = {Corteri Intelligence}
}

Acknowledgements

Distilled from amazon/chronos-2. Evaluated with GIFT-Eval. Pretraining corpora: LOTSA, Time-300B, GIFT-Eval-Pretrain.