Next-Hour Volatility Forecast

Forecast next-hour realized volatility from trailing returns

Pro

The dataset contains, for each of ten anonymous tokens, trailing one-minute log returns and trading volume observed at hourly forecast points. The task is to forecast each token's realized volatility over the following hour, computed from one-minute returns.

Training and test data are split chronologically with an embargo gap between them.

Prediction target

Realized volatility over the next hour

\sigma \;=\; \sqrt{\sum_{k=1}^{M} r_k^{2}}, \qquad r_k = \ln\!\left(\dfrac{p_k}{p_{k-1}}\right)

where r_1 ... r_M are the token's one-minute log returns over the next hour (about 60 of them) and p_k are its one-minute prices.

Scoring

Submissions are graded against the hidden holdout by root mean squared error in log space, lower is better:

\mathrm{RMSE}_{\log} \;=\; \sqrt{\dfrac{1}{N}\sum_{i=1}^{N}\left(\ln \hat{y}_i - \ln y_i\right)^2}

ŷᵢ is your predicted volatility and yᵢ the realized volatility for each of the N scored rows. Log space makes the error a ratio, so being off by a given factor costs the same whether volatility is high or low.

Data files

train.csv.gz and test.csv.gz. The test set has 21,432 rows; your submission must cover every test id exactly once.

Data dictionary

Column	Description
`id`	Opaque row id. Your submission joins on this; row order does not matter.
`window_start`	Integer minutes since the start of the data. The forecast is made at this moment; lag features end here.
`window_end`	window_start + 60. The target is the realized volatility over this window. Group across time however you like; validate forward in time.
`asset`	Stable anonymous token id (0-9). The same asset keeps the same id across all rows.
`lag_m`	A trailing one-minute log return of the token: lag_m covers the minute ending m minutes before window_start, so lag_m = ln( price at (window_start - (m-1) min) / price at (window_start - m min) ). The offsets m present in the header are dense for recent minutes and sparser further back; lag_1 is the most recent minute.
`volume_m`	The token's trading volume during the same minute that lag_m covers (base-asset volume of that one-minute bar), at the same offsets m as the return lags. Volume and volatility move together, so this is a useful predictor.
`target`	Realized volatility over the next hour, computed from 1-minute returns (train only).

Rules

5 scored submissions per day (UTC).
Your submission is scored on a public split (shown to you) and a private split (used for ranking). Your leaderboard entry is the private score of your best public submission.
The board is always open. No deadlines, all-time rankings.
Target windows never overlap for the same asset, and the test range is strictly later than training with an embargo gap between them.

Getting started

import pandas as pd

train = pd.read_csv('train.csv.gz')
test = pd.read_csv('test.csv.gz')
lags = [c for c in train.columns if c.startswith('lag_')]
# Baseline: root mean of recent squared returns as the vol proxy
test['prediction'] = (test[lags[:12]] ** 2).mean(axis=1) ** 0.5
test[['id', 'prediction']].to_csv('submission.csv', index=False)