Developers
Build and Deploy a Forecaster

Build and Deploy a Forecaster

The Allora Forecaster is designed to run a model that predicts how accurate inferers are at arbitrary tasks. Any forecaster can be augmented using proprietary data sources, which likely overlap with the data used by inference models. A boilerplate forecaster (opens in a new tab) has been provided that has demonstrated ability for arbitrary topics.

Forecaster Components Overview

ComponentPurposeKey Functions
Data IndexingRetrieves necessary data from the blockchain using the Postgres indexer.Utilizes the extract folder for querying data from Postgres and making it accessible to the forecaster.
ModelingCore functionality for model selection and training.Supports different machine learning algorithms like LightGBM and XGBoost.
Prediction EngineRuns selected models on historical data to generate future predictions.Ingests time-series data and outputs forecast values based on the chosen model.
Model PlotsVisualizes model performance and forecast accuracy.Generates plots such as Prediction vs Actual, Residuals, and Forecast Horizon for intuitive evaluation.
Performance MetricsMeasures the accuracy and effectiveness of model predictions.Key metrics include MAE (Mean Absolute Error), RMSE (Root Mean Squared Error), R2 Score, Mean Absolute Percentage Error, Median Absolute Percentage Error
Scoring MechanismAssigns scores based on model performance compared to other participants.Determines which forecasts contribute to the Allora Network’s final consensus based on accuracy and uniqueness.