How BaseModel Works

BaseModel uses a two-stage approach to build predictive models from your event data.

The Two Stages

Benefits

1. Efficiency — Train the foundation once, reuse for unlimited downstream tasks
2. Accuracy — The foundation learns rich behavioral patterns that benefit all predictions
3. Consistency — All downstream models share the same learned representations
4. Speed — New prediction tasks take dramatically less time than building from scratch

Inputs: Behavioral Data & Attributes

Unlike traditional ML that requires hand-crafted features, BaseModel learns directly from behavioral data, i.e. timestamped, entity-bound events such as:

• Purchase histories
• Click streams
• Session logs
• Subscription events
• Support history etc.

In addition to events, BaseModel ingests entity attributes — static or slowly changing properties such as demographics, account type, or region — to enrich the behavioral context.

The foundation model combines these raw event sequences and attributes, learning meaningful patterns on its own — no feature engineering required.

Stage 1: Training Foundation Model

The foundation model builds the understanding of how entities (customers, users, accounts) behave over time. See below for an intuition of what it learns:

These patterns are captured automatically and encoded into dense entity representations that power all downstream tasks.

Stage 2: Fine-Tuning for Prediction Tasks

The Timeline Split

During training, past data is time-split into:

• History — Events before the split point, used as model input
• Future — Events after the split point, used to generate the prediction target

The split point moves randomly during training for robust generalization.

Target for Transfer Learning

Once the foundation model is trained, you define a target function — a simple Python function that specifies what to predict.

def churn_target(history, future, attributes, ctx):
    """
    history  = events before the split (model input)
    future   = events after the split (used to generate labels)
    """
    if len(future["transactions"]) == 0:
        return 1  # Churned
    return 0      # Active

The foundation model handles everything else — no additional feature work, no retraining from scratch.