Guides / 31 October 2026

A NIFTY options dataset for machine learning: features, labels and leak-free splits

Using NIFTY options history as a machine learning dataset: turning minute bars into features, labelling without look-ahead, and splitting by time and expiry.

Market data makes it easy to build a model that looks brilliant and fails live. Most failures are data mistakes: features that peek at the future, or train and test sets that share a session. This guide sets up a NIFTY options dataset so those mistakes are hard to make.

Features from minute bars

import pandas as pd

df = pd.read_parquet("nifty_atm_straddle_1min.parquet")   # one row per minute
df = df.set_index("datetime").sort_index()
f = pd.DataFrame(index=df.index)
f["ret_5"] = df["straddle"].pct_change(5)
f["ret_15"] = df["straddle"].pct_change(15)
f["vol_15"] = df["straddle"].pct_change().rolling(15).std()
f["oi_chg"] = df["oi"].diff(15)
f["mins_to_close"] = (15 * 60 + 30) - (df.index.hour * 60 + df.index.minute)

Everything above uses only past and current bars.

Labels without look-ahead

Define the label as the change over the next N minutes, then drop rows where those N minutes cross the end of the session. Never fill a label from the next day.

N = 15
f["y"] = df["straddle"].shift(-N) / df["straddle"] - 1
same_day = df.index.normalize() == df.index.to_series().shift(-N).dt.normalize()
f = f[same_day].dropna()

Splitting

• Split by date, never by random rows: minutes in one session are strongly related.
• Leave at least one full day out between train and test.
• Keep whole expiry weeks together; a weekly series should sit entirely on one side.

More history gives more independent sessions: 3 years or since 2020.