🚚 Chile Spatial Logistics Optimizer

Geospatial intelligence system for Dark Store siting and last-mile delivery route optimization in Santiago, Chile's Metropolitan Region.

Code and method Full README

This page shows the project's results. The methodology, the design decisions and the limitations are documented in the repository's README.

6.1 Demand forecasting (Phase 3) -- model comparison

Model comparison — The animated version below traces the loss curves epoch by epoch, with a floating label tracking each activation's current loss.

Activation convergence animated — Honest note on forecasting accuracy: R² near zero across all five models is the correct, un-massaged result here — total_demand in the synthetic generator is drawn independently of location (rng.randint(min_demand, max_demand) per order, uncorrelated with comuna or coordinates by construction), so there is little real spatial signal to learn from location alone. The comparison is still meaningful as a methodology demonstration — three genuinely different model families, a domain-specific asymmetric loss, and an honest activation comparison — that would show a real accuracy gap on demand data with actual spatial structure (seasonality, comuna-level income/population correlates, proximity to transit), which this synthetic generator doesn't model.

Activation convergence — Honest note on forecasting accuracy: R² near zero across all five models is the correct, un-massaged result here — total_demand in the synthetic generator is drawn independently of location (rng.randint(min_demand, max_demand) per order, uncorrelated with comuna or coordinates by construction), so there is little real spatial signal to learn from location alone. The comparison is still meaningful as a methodology demonstration — three genuinely different model families, a domain-specific asymmetric loss, and an honest activation comparison — that would show a real accuracy gap on demand data with actual spatial structure (seasonality, comuna-level income/population correlates, proximity to transit), which this synthetic generator doesn't model.

Actual vs predicted — Honest note on forecasting accuracy: R² near zero across all five models is the correct, un-massaged result here — total_demand in the synthetic generator is drawn independently of location (rng.randint(min_demand, max_demand) per order, uncorrelated with comuna or coordinates by construction), so there is little real spatial signal to learn from location alone. The comparison is still meaningful as a methodology demonstration — three genuinely different model families, a domain-specific asymmetric loss, and an honest activation comparison — that would show a real accuracy gap on demand data with actual spatial structure (seasonality, comuna-level income/population correlates, proximity to transit), which this synthetic generator doesn't model.