NanoForecast v0.5: Competitive Time Series Forecasting Through Training Pipeline Optimization
Gautam Kishore
Eulogik
Published September 15, 2026 · on the feed September 29, 2026 · ▲ 2 · arXiv 2609.31669
The abstract
We present NanoForecast v0.5, a 6.5M-parameter forecaster that competes with models 31x its size (TimesFM, 200M parameters) after training pipeline fixes and no architecture change. Retraining the v0.3 architecture with corrected loss-scope handling, tensor shape alignment, and wider augmentation coverage cuts overall Mean Absolute Scaled Error by 43.8% under one fixed protocol (MASE 3.030 to 1.704) on the same data and compute budget. NanoForecast v0.5 beats TimesFM on all three ETT datasets (MASE 0.676/1.110/0.287 vs. 0.705/1.360/0.545) and on exchange rate (4.317 vs. 4.383); TimesFM keeps a clear lead on the high-cardinality electricity and traffic sets. Against PatchTST (15M+ parameters, official configuration), v0.5 wins all three ETT sets. Training takes about 12 hours on a single cloud GPU (NVIDIA T4, Google Colab) and inference needs no GPU (measurements in this paper are on an Apple M4 CPU). We release all code, pretrained checkpoints, and evaluation framework under Apache 2.0 at https://github.com/eulogik/NanoForecast
The Brief
2:00A short podcast summary of the paper. Written and voiced by AI.
Transcript
NanoForecast v0.5. Competitive Time Series Forecasting Through Training Pipeline Optimization. By Gautam Kishore, from Eulogik. Published on the fifteenth of September. Its code repository is public and real: about thirty source files.
The claim is that fixing the training pipeline, with no change to the architecture, was enough to make a very small forecaster competitive with one thirty-one times its size. NanoForecast v0.5 has six point five million parameters. The comparison model is TimesFM, at two hundred million.
The v0.3 architecture was retrained with three corrections: loss-scope handling, tensor shape alignment, and wider augmentation coverage. On the same data and compute budget, under one fixed protocol, overall Mean Absolute Scaled Error fell by forty-three point eight percent, from three point zero three zero to one point seven zero four.
Against TimesFM, v0.5 wins on all three ETT datasets. The widest of those margins is an error of zero point two eight seven against TimesFM's zero point five four five. It also edges ahead on exchange rate, four point three one seven against four point three eight three. Against PatchTST, run in its official configuration at fifteen million parameters or more, v0.5 wins all three ETT sets.
The win is not universal. TimesFM keeps a clear lead on the high-cardinality electricity and traffic sets. The forty-three point eight percent improvement is stated under one fixed protocol. Training takes about twelve hours on a single cloud GPU, an NVIDIA T4 on Google Colab, inference needs no GPU, and the paper's measurements were made on an Apple M4 CPU. Code, pretrained checkpoints and the evaluation framework are released under Apache 2.0.
Nothing in the model changed between versions. What changed was how it was trained, and that alone cut its error nearly in half and put it ahead of a model thirty-one times larger on four datasets, while the high-cardinality sets stayed with the bigger model.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A tiny model, a giant rival, and training curves that look perfectly fine.
You built a small forecaster, 6.5M parameters, light enough to run without a GPU. Its overall error is poor, and the 200M-parameter rival looks out of reach.
The road ahead
- 1 · The problem ~6 min A tiny forecaster trains without complaint yet scores badly, and the easy story is that small models just cannot keep up.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Three silent bugs in training, not the model's design, held a small forecaster back; fix them and retrain the same model.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Keep the design fixed, fix three pipeline bugs, retrain on the same data and budget, and score every model one way.
- 6 · The experiment ~6 min You’ll guess: With training pipeline fixes alone and no architecture change, by how much did NanoForecast's overall Mean Absolute Scaled Error drop on the same data and compute?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.