FAQ: Machine Learning Platform

On which platform does your Large Language Model (LLM) sit?: Our Demand Forecasting solution does not utilize LLMs, as Large Language Models are designed for natural language processing rather than direct numerical prediction. Instead, we offer two specialized forecasting models optimized for accuracy and cost-efficiency:
Trend-based Forecasting (Default): This model uses time-series forecasting with advanced weighting and intelligence. It determines the most likely outcomes by analyzing current trends against historical patterns (Example: comparing current performance to similar days from the previous year). This is ideal for metrics with regular, seasonal patterns.
Gradient-boosted Tree Model: This machine learning framework utilizes multiple decision trees to categorize data based on specific criteria. It is designed to capture complex, non-linear relationships between various factors that influence demand.

To ensure maximum precision, each of our models is trained in isolation. This means a unique model is built specifically for each individual metric to understand its distinct patterns; insights gained from one metric are never applied to another. While LLMs are powerful for text, they require data to be converted into natural language descriptions, which typically results in lower accuracy and higher costs for time-series forecasting compared to our specialized numerical models.
How do you mitigate backflow and leakage?: It's important to clarify the architectural safeguards of our forecasting environment. Our system is designed around the principles of isolation and ephemerality:
Model Isolation (No Cross-Pollination)
Our Gradient-boosted Tree and Trend-based models are trained on a per-customer, per-metric basis. We do not use global models or Large Language Models (LLMs) that pool data from multiple clients. Your data remains in a silo; there is no mathematical mechanism for your business patterns to influence the forecasts of another customer, or vice versa. Additionally, a model trained today will know nothing about the model trained a week ago for that same metric, effectively isolating it from itself. This is also why our system requires 1 - 3 years of historical actuals, because as we retrain the new model has never seen those values before and must learn from scratch.
Stateless Retraining (Zero Persistence) - Every forecasting run involves a complete retraining of the model from scratch.
No Memory: We do not maintain a persistent model that retains information across sessions.
Trend-based Forecasting Model: Once the forecast is generated, the specific model instance is discarded. Because the model is 'reborn' for every run using only the currently provided data, there is no risk of residual data 'back-flowing' into future iterations.
Gradient-boosted Tree Model: We back up the specific model generated during a forecasting run, however this is for troubleshooting, learning, and support and is not used again in the next forecasting run.