Apply Machine Learning

Show the modeling path immediately after the data, just like the best quant product pages do.

The goal here is to keep the product narrative concrete. Users should not have to infer what happens between downloading data and generating predictions.

import lightgbm as lgb
import pandas as pd

training_data = pd.read_parquet("train.parquet")
features = [c for c in training_data.columns if "feature" in c]

model = lgb.LGBMRegressor(
    n_estimators=2000,
    learning_rate=0.01,
    max_depth=5,
    num_leaves=32,
    colsample_bytree=0.1,
)

model.fit(
    training_data[features],
    training_data["target"],
)

live_data = pd.read_parquet("live.parquet")
predictions = model.predict(live_data[features])

What this page is doing

Numer.ai uses code snippets to shorten the path from curiosity to action. This page expands that approach into a dedicated modeling page where code, intent, and workflow sit together.

Clarify how features are selected.
Expose a default estimator and training setup.
Bridge naturally from training to live inference.
Set up the next page on submissions and scoring.

Workflow

Modeling is not a black box. It is a sequence that should be teachable in one pass.

1. Prepare Features

Isolate train and live features under one shared schema.

2. Train A Baseline

Start with a strong baseline before chasing exotic architectures.

3. Generate Live Predictions

Push the output toward a scored submission instead of stopping at offline metrics.