Data

Latest version

Signals v3.0 was released in August 2026. It adds the Jupiter targets, expands the neutralizer from 200 to 300 columns, and removes sample-weight files.

Data Structure

The Numerai Signals dataset is a tabular dataset that describes the global stock market over time.

At a high level, each row represents a stock at a specific point in time identified by its numerai_ticker or composite_figi and the date. The date represents the day that market close data was used to generate the features - in the context of rounds and live data, this is the day before a round opens. The targets measure future returns over horizons such as 20 or 60 days relative to the date.

Features

The Signals dataset contains:

1. Factors, which are similar to the factors which the target is neutral to. Use these to determine if your model is sufficiently unique, use them to neutralize your predictions, or use them as additional features in your dataset.
2. Some starter features, which are relatively simple classic quant features constructed from returns series.

Features

The starter features in Signals v3.0 are:

feature_adv_20d_factor
feature_beta_factor
feature_book_to_price_factor
feature_country
feature_dividend_yield_factor
feature_earnings_yield_factor
feature_growth_factor
feature_impact_cost_factor
feature_market_cap_factor
feature_momentum_12w_factor
feature_momentum_26w_factor
feature_momentum_52w_factor
feature_momentum_52w_less_4w_factor
feature_ppo_60d_130d_country_ranknorm
feature_ppo_60d_90d_country_ranknorm
feature_price_factor
feature_rsi_130d_country_ranknorm
feature_rsi_60d_country_ranknorm
feature_rsi_90d_country_ranknorm
feature_trix_130d_country_ranknorm
feature_trix_60d_country_ranknorm
feature_value_factor
feature_volatility_factor

Features with {n}(d|w) in the name (for example, feature_adv_20d_factor) are time-series features that are computed over n days or n weeks.

Features with country_ranknorm in the name are grouped by country, then ranked, then gaussianized.

Features with factor in the name refer to risk factors that most of the targets are neutral to. Price factor is grouped by country, then ranked.

PPO is a percentage price oscillator that compares shorter and longer moving averages in a ratio.

RSI is the relative strength index usually used as an overbought/oversold indicator.

TRIX is a triple exponential moving average indicator usually used as a momentum or reversal feature.

momentum_52w_less_4w refers to the one-year return of a stock excluding the last 4 weeks.

Targets

The target of the dataset is specifically engineered to match the strategy of the hedge fund.

Given our hedge fund is market/country/sector and factor neutral, you can basically interpret the target as stock-specific returns that are not explained by broader trends in the market/country/sector or well-known factors. In simple terms: what we are after is "alpha".

V3 Targets

Signals v3.0 adds two Jupiter targets:

• target_jupiter_20 uses a 20-day horizon.
• target_jupiter_60 uses a 60-day horizon and is the primary v3 target.

The target column is an exact alias for target_jupiter_60.

Other Targets

Signals also provides targets for research, comparison, and models that use a different return horizon or neutralization:

• target_raw_return_20 and target_raw_return_60 account for return adjustments such as stock splits and dividends, then bin the returns into Numerai's standard distribution.
• target_factor_neutral_20 and target_factor_neutral_60 are neutralized to standard factors such as country, sector, beta, momentum, and size.
• target_factor_feat_neutral_20 and target_factor_feat_neutral_60 are neutralized to a larger set of factors and features.

Auxiliary Targets

Apart from the main target we provide many auxiliary targets that are different types of stock-specific returns. Like the main target, these auxiliary targets are also based on stock-specific returns but are different in what is residualized (eg. market/country vs sector/factor) and time horizon (eg. 20 days vs 60 days).

Even though our objective is to predict the main target, we have found it helpful to also model these auxiliary targets. Sometimes, a model trained on an auxiliary target can even outperform a model trained on the main target. In other scenarios, we have found that building an ensemble of models trained on different targets can also help with performance.

Some target values can be NaN when their return horizon has not finished. Recent validation rows can have a null target_jupiter_60 and target; filter out null values for the target you use.

Data API

The best way to access the Signals dataset is via the data API.

Files

The Numerai dataset is made up of many different files.

Here is how to query the data API to see what files are available and how to download a file.

from numerapi import SignalsAPI
napi = SignalsAPI()

DATA_VERSION = "signals/v3.0"

[f for f in napi.list_datasets() if f.startswith(DATA_VERSION)] 

[f'{DATA_VERSION}/live.parquet',
 f'{DATA_VERSION}/live_example_preds.csv',
 f'{DATA_VERSION}/live_example_preds.parquet',
 f'{DATA_VERSION}/live_neutralizer.parquet',
 f'{DATA_VERSION}/train.parquet',
 f'{DATA_VERSION}/train_neutralizer.parquet',
 f'{DATA_VERSION}/validation.parquet',
 f'{DATA_VERSION}/validation_example_preds.csv',
 f'{DATA_VERSION}/validation_example_preds.parquet',
 f'{DATA_VERSION}/validation_neutralizer.parquet']
 
# Download the training data 
napi.download_dataset(f'{DATA_VERSION}/train.parquet')

• train.parquet contains historical tickers, features, and targets for training.
• validation.parquet contains later historical tickers, features, and targets for validation. It expands weekly, and recent rows can have null 60-day targets until their returns mature.
• live.parquet contains the current live universe and features with no realized targets.
• train_neutralizer.parquet, validation_neutralizer.parquet, and live_neutralizer.parquet contain the matching 300-column neutralizer matrices.
• validation_example_preds and live_example_preds contain example predictions in CSV and Parquet formats.

Signals v3.0 does not include sample-weight files. Signals v2.1 remains available for older pipelines that still require them.