IndiaVol — Chronos-2 fine-tuned for NSE realized volatility

IndiaVol forecasts how volatile Indian stocks and indices will be over the next 1, 5 and 10 trading days. It is amazon/chronos-2, fully fine-tuned on daily realized variance of 5 NSE/BSE indices and ~206 NSE F&O stocks (2022-01 → 2024-06).

It does not predict price direction, and it is not a trading strategy. It is a volatility forecast, useful as an input for option pricing, position sizing and stops.

Code, pre-registration and full evaluation: https://github.com/chelvadataer/india-vol

Headline result (sealed holdout, scored once)

The evaluation plan and publish rule were written down before any model was run (pre-registration). The holdout (2025-01 → 2026-10, 211 series, 20,323 forecast origins) was opened once, after the model variant and its bias correction had been frozen on the validation period.

QLIKE on variance (lower is better):

model	h = 1	h = 5	h = 10
RV20 (trailing 20-day close-to-close vol)	0.562	0.351	0.306
HAR (log-HAR, direct h-step, expanding OLS)	0.317	0.190	0.170
Chronos-2 zero-shot (bias-corrected)	0.212	0.134	0.122
IndiaVol (this model, bias-corrected)	0.205	0.130	0.117

Publish rule (h = 5): beat HAR in a pooled day-block bootstrap and on ≥ 4 of 5 indices.

Pooled QLIKE(IndiaVol) − QLIKE(HAR) = −0.060, 95 % CI [−0.077, −0.044].

Per index (Diebold-Mariano, HAC, one-sided):

index	IndiaVol	HAR	DM t	p
NIFTY	0.176	0.246	−2.75	0.003
BANKNIFTY	0.166	0.239	−2.63	0.004
FINNIFTY	0.165	0.232	−2.82	0.002
MIDCPNIFTY	0.158	0.193	−2.70	0.004
SENSEX	0.178	0.247	−2.68	0.004

Rule passed: 5 / 5.

Read this before you use it

Most of the gain over HAR comes from Chronos-2 itself, not from the fine-tune. On the holdout, fine-tuning adds a small further improvement over zero-shot Chronos-2 (h = 5: −0.004, 95 % CI [−0.009, −0.0002]; a secondary, not pre-registered comparison). On the validation period zero-shot was slightly better (0.1287 vs 0.1301). If you only need "better than HAR", bias-corrected zero-shot Chronos-2 gets you almost all of the way.
Bias correction is part of the model. Raw medians are biased; apply the frozen factors below.
Pretraining overlap: we cannot rule out that Chronos-2's own pretraining corpus contains series overlapping the 2025 holdout period.
Survivorship: the stock universe is today's F&O list, not point-in-time.
Not trading advice. A better volatility forecast does not by itself make money.

Training details


Target series	daily log realized variance: Σ squared 5-min log returns 09:15–15:30 IST + squared overnight return
Data	1-minute bars, public Upstox v3 historical-candle API (not redistributed)
Train / val / holdout	≤ 2024-06-30 / 2024-07 → 2024-12 / 2025-01 → 2026-10
Candidates (picked on validation h = 5 QLIKE)	LoRA lr 1e-5 × 1,000 and × 3,000 steps; full fine-tune lr 1e-6 × 1,000 steps (selected)
Context / horizon	512 days / 10 days, batch 64, CPU
Forecast origins	indices every trading day, stocks every 5th trading day

Usage

import math, numpy as np, torch
from chronos import Chronos2Pipeline   # pip install "chronos-forecasting>=2.3"

pipe = Chronos2Pipeline.from_pretrained("chelvaser/indiavol-chronos2", device_map="cpu")
BIAS = {1: 0.016230, 5: -0.079483, 10: -0.103922}   # frozen on validation (see indiavol_config.json)

# log_rv: 1-D array of daily log realized variance up to and including today (most recent last)
q, _ = pipe.predict_quantiles([torch.tensor(log_rv[-512:], dtype=torch.float32)],
                              prediction_length=10, quantile_levels=[0.5])
med = q[0][0, :, 0].numpy() if q[0].dim() == 3 else q[0][:, 0].numpy()
sigma = {h: math.sqrt(252 * np.exp(med[:h]).mean()) * math.exp(-BIAS[h]) for h in (1, 5, 10)}
print(sigma)   # annualised σ forecast over the next 1 / 5 / 10 trading days