Kronos-NSE-1d — a fine-tune that improved calibration and produced no alpha

Kronos-small and Kronos-Tokenizer-base fine-tuned on NSE (India) daily equity bars, evaluated for one-day-ahead cross-sectional ranking skill.

This model does not predict Indian equity returns. It is published as a documented negative result together with a real, measured calibration improvement. Do not trade it.

Headline

	pretrained Kronos-small	this fine-tune
Rank-IC (1d, 100 sessions)	−0.0116	−0.0114
IC t-statistic	−0.54	−0.50
Directional hit rate	50.25%	51.57%
Forecast dispersion ÷ realised	3.61×	1.85×
Forecasts implying >±10% in a day	11.82%	2.93%
Long/short decile spread	0.046%	0.139%

Fine-tuning roughly halved forecast dispersion and cut impossible one-day forecasts by 75%. It did not create ranking skill: the information coefficient is statistically indistinguishable from zero, and marginally negative.

For scale, the decile spread of 0.139% sits an order of magnitude below the ~10 bps round-trip cost of trading NSE equities, so it is not tradeable even if it were real.

The baseline beat it

A 12-1 momentum signal, measured on the identical 100 sessions at the identical horizon:

signal	rank-IC	t	permutation control
12-1 momentum	+0.0376	+3.51	+0.0003
this model	−0.0114	−0.50	—

A decades-old, six-line ranking rule outperforms the fine-tuned foundation model by a wide and statistically significant margin. That comparison is the most useful thing in this card.

(Momentum's own skill decays: +0.0596 over 2015-2019 versus +0.0290 over 2023-2026. And in a cost-aware backtest over 2023-2026 a concentrated top-10 momentum book returned 7.78% CAGR net against 20.46% for simply holding the universe equal-weighted. A positive IC is necessary for a tradeable strategy, not sufficient.)

Evaluation

• Held-out window: 2026-01-01 → 2026-07-30, 100 sessions, never seen in training.
• Universe: top 80 by 60-day median turnover, re-formed each session.
• Observations: 7,465 forecasts per variant.
• Metric: cross-sectional Spearman rank-IC of predicted vs realised next-session close-to-close return, aggregated as a t-statistic over sessions.
• Sampling: 1-day horizon on daily dates, so windows do not overlap and the t-statistic is not inflated.

Four variants were evaluated. None showed skill:

variant	rank-IC	t	dispersion	impossible
pretrained Kronos-small	−0.0116	−0.54	3.61×	11.82%
fine-tuned, 4y raw (this model)	−0.0114	−0.50	1.85×	2.93%
fine-tuned, 15y corporate-action-adjusted	+0.0021	+0.09	10.19×	56.61%
fine-tuned, 15y, frozen tokenizer	−0.0315	−2.05	4.85×	17.86%

The 15-year variant, despite four times the data and corporate-action adjustment, regressed calibration badly. Freezing the tokenizer recovered about half of that, implicating tokenizer/predictor mismatch as a partial cause — the tokenizer received far more adaptation than the predictor.

The frozen-tokenizer arm's t of −2.05 is the only |t| above 2 in the study. It should not be believed: it is the fourth of four variants tested, where a |t| > 2 arises by chance roughly 18% of the time, and no prior hypothesis predicted negative skill.

Training

• Base: NeoQuasar/Kronos-small (predictor), NeoQuasar/Kronos-Tokenizer-base
• Data: NSE daily bars, ~300 symbols by liquidity, 2022-09 → 2025-06 targets
• Windows: 174,645 (90-bar lookback, 10-bar horizon)
• Schedule: tokenizer 3 epochs, then predictor against it; AdamW, lr 2e-4 / 4e-5, batch 16, grad-clip 3.0, MPS
• Checkpoint selection: best validation loss. The predictor's validation loss turned after epoch 1 and epochs 2-3 were discarded — on a corpus five times larger the same thing happened, so this is not a data-volume limitation.
• Normalisation: statistics from the lookback segment only, never the predicted segment.

Usage

from huggingface_hub import snapshot_download
from model import Kronos, KronosTokenizer, KronosPredictor  # upstream Kronos repo

path = snapshot_download("Sahilsingh0808/kronos-nse-1d")
tokenizer = KronosTokenizer.from_pretrained(f"{path}/tokenizer").eval()
predictor = Kronos.from_pretrained(f"{path}/predictor").eval()
kp = KronosPredictor(predictor, tokenizer, max_context=512)

.eval() matters: from_pretrained leaves these modules in train mode, where a non-zero attention dropout reaches scaled_dot_product_attention — an outright error on MPS, and silent noise in every forecast elsewhere.

Sample at temperature 0.6, not 1.0. At 1.0 the pretrained model's forecast spread measured 3.95× realised volatility with 12.6% impossible forecasts.

Intended use

Research into financial time-series foundation models; a calibration reference; a reproducible negative result for anyone evaluating whether to fine-tune Kronos on a new equity market.

Out of scope

Trading, investment decisions, or any use where the output is treated as a return forecast. The model has no measured predictive skill. It is not investment advice.

Limitations

• Evaluated on one market (NSE), one horizon (1 day), one window (100 sessions of 2026). 100 sessions is enough to reject a large effect, not to rule out a small one — detecting an IC of 0.02 at |t| = 2 would need roughly 400 sessions.
• The training corpus was not corporate-action adjusted, so split and bonus gaps appear as price crashes. The adjusted 15-year variant fixed that and performed worse, so this is not the binding constraint, but it is a real defect in this checkpoint.
• Training universe is liquidity-ranked, biased toward large caps.
• One market regime: 2022-2026 Indian equities were broadly rising.

Citation

@misc{kronos_nse_1d_2026,
  title  = {Kronos-NSE-1d: fine-tuning a financial foundation model on Indian
            equities improves calibration but not predictive skill},
  author = {Sahil Singh},
  year   = {2026},
  note   = {Negative result. Rank-IC -0.0114 (t -0.50) over 100 held-out
            sessions; a 12-1 momentum baseline scored +0.0376 (t +3.51) on
            identical terms.},
  url    = {https://huggingface.co/Sahilsingh0808/kronos-nse-1d}
}