LT
LTprophecy
Docs/Core Features/Forecasting

Forecasting

SalesForecaster's forecasting engine combines multiple ML models with AI-discovered features to produce accurate, explainable predictions.

How Forecasting Works

The Ensemble Approach

Rather than relying on a single model, SalesForecaster trains an ensemble of diverse algorithms:

  • XGBoost — Gradient-boosted trees excelling at tabular data
  • LightGBM — Fast gradient boosting with categorical feature support
  • Neural Networks — Deep learning models for complex non-linear patterns
  • Statistical Models — ARIMA and Prophet for trend and seasonality decomposition

The ensemble combines predictions using a learned weighting scheme, producing more robust forecasts than any individual model.

Feature Enrichment

Each forecast is powered by two categories of features:

  1. Endogenous features — Derived from your sales data (lags, trends, seasonality)
  2. Exogenous features — Discovered by AI research agents (market signals, economic indicators, competitor actions)

Creating a Forecast

Quick Forecast

For rapid predictions without full model training:

  1. Go to Forecasts > New Forecast
  2. Select a dataset
  3. Choose Quick Forecast
  4. Set the forecast horizon
  5. Click Generate

Quick forecasts use pre-trained model templates and typically complete in under 30 seconds.

Full Forecast

For maximum accuracy:

  1. Train a model first under Models > Train
  2. Navigate to Forecasts > New Forecast
  3. Select your trained model
  4. Configure the forecast horizon and parameters
  5. Click Generate

Forecast Output

Every forecast includes:

Point Predictions

The primary forecast values for each future time period, accompanied by prediction intervals at 80% and 95% confidence levels.

Feature Importance

A breakdown of which factors contributed most to each prediction, including:

  • Relative importance scores
  • Directional impact (positive or negative)
  • Comparison to baseline expectations

Scenario Analysis

Compare your forecast against alternative scenarios:

  • Optimistic — Best-case assumptions for key drivers
  • Pessimistic — Downside risk modeling
  • Custom — Define your own scenario parameters

Scheduled Forecasts

Automate recurring forecasts:

  1. Go to Forecasts > Scheduled
  2. Click New Schedule
  3. Configure frequency (daily, weekly, monthly)
  4. Select the model and parameters
  5. Set up notification preferences

Scheduled forecasts automatically incorporate the latest data and alert you to significant changes.

Model Monitoring

After deploying a forecast, SalesForecaster continuously monitors performance:

  • Prediction accuracy — Actual vs. predicted comparisons
  • Model drift — Detects when model accuracy degrades
  • Data drift — Alerts when input data distributions shift
  • Retraining triggers — Automatically suggests or initiates retraining
Datasets