lawrencehui/chronos-2-empty-history-fix

Chronos-2, empty-history fix

This is amazon/chronos-2 with 1,536 of its 120 million weights changed, and no retraining. The change fixes one bug: when the history you give it is empty, Chronos-2 returns huge forecast ranges.

History given to the model	Original 10%–90% range	This model
No data at all (all NaN)	−141,915 to +130,875	−0.4 to 0.2
One real value (7.0)	5.2 to 8.9	7.0 to 7.0
Normal series around 10	9.3 to 10.6	9.3 to 10.6, bit-identical

With two or more real values in the history, the output is bit-for-bit identical to the original.

Use

Load it the same way as the original:

from chronos import Chronos2Pipeline
import numpy as np

pipe = Chronos2Pipeline.from_pretrained("lawrencehui/chronos-2-empty-history-fix", device_map="cpu")
pipe.predict_quantiles([{"target": np.full(500, np.nan, dtype=np.float32)}],
                       prediction_length=24, quantile_levels=[0.1, 0.5, 0.9])

What changed

One neuron in the last encoder block's MLP never activated. An AI agent (Claude Opus 5.5, in Claude Code) gave it a job, working from the weights and practice data, with 50 test runs allowed and no gradients. It used 11 runs and about 30 minutes.

• encoder.block.11 MLP wi, row 11 (768 weights): detects an empty history
• encoder.block.11 MLP wo, column 11 (768 weights): steers that forecast to a flat line near zero

The 8 KB delta, the agent's full notebook and all the evaluation code are in github.com/lawrencehui/Chronos-2-weight-space-fix.

Results on GIFT-Eval (test split)

Only 12 of the 97 tasks contain a series with at most one real value in its history, so only those can change. All of them are Bitbrains (cloud server logs), electricity/W or kdd_cup_2018.

Task	Original	Fixed
bitbrains_rnd / 5T / medium	1.0046	0.6700 (−33.3%)
bitbrains_rnd / 5T / long	0.8730	0.6254 (−28.4%)
bitbrains_rnd / H / short	0.8035	0.6702 (−16.6%)
5 other Bitbrains tasks		0.2–3.9% better
Every other task		unchanged
Overall normalised CRPS	0.4854	0.4803

Point forecasts (MASE) are unchanged at 0.6978. On our computation that moves Chronos-2 from #79 to #69 of 164 on the leaderboard. This is not an official leaderboard submission.

Caveats

• A one-line code check ("empty history → flat forecast") fixes the same bug, and that's the fix we suggested upstream in amazon-science/chronos-forecasting#544. Prefer that if it lands. This model exists to show that the fix can live in the weights, and so you can try it.
• With exactly one real value, the forecast collapses onto that value (zero width). That's no worse than the original here, but it's not a calibrated range.
• Tested with chronos-forecasting 2.3.2 on CPU (Apple Silicon).

Credits

• Base model: amazon/chronos-2 by Amazon, Apache-2.0. All credit for the model is theirs; this repo only changes 1,536 weights.
• Method: Sparks of Weight-Space Programming by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab.
• Write-up: lawrenceh.xyz/writings/chronos-2-one-dead-neuron