Sendy Logistics — Delivery Time Prediction

The Problem

Predicting delivery duration in Nairobi sounds straightforward until you realize how many factors a simple formula misses. A motorcycle courier picking up a package has a known origin, a known destination, and a calculable distance. You could estimate duration as distance divided by average speed. But Nairobi traffic is wildly variable. Rider experience matters. Pickup delays matter. The time of day, the neighborhood, and the day of week all interact in ways that a formula cannot capture.

The competition provided historical delivery records from Sendy, a logistics platform, with timestamps for placement, confirmation, pickup, and arrival. The goal was to predict the total duration from pickup to arrival. The evaluation metric was RMSE, meaning large errors are penalized disproportionately — getting a few deliveries very wrong hurts more than being slightly off on many.

The interesting tension in this problem is between the physics baseline (distance/speed) and the contextual complexity (traffic, rider behavior, location-specific delays). Neither alone gives a good answer. The formula is too simple. Pure ML ignores obvious domain knowledge. The question becomes: how do you combine both?

My Approach

The first thing I did was calculate what a formula would give: distance divided by average rider speed. Then I asked the model to learn the residual — how much the real world deviates from this simple estimate, and in which direction. Rather than using the formula as the prediction or ignoring it entirely, I fed the analytical estimate into the model as a feature. The model's task became: predict how wrong the baseline estimate is.

This is a powerful pattern. The domain formula captures the structural component (longer distances take longer), while the ML model captures the contextual corrections (this rider is faster, this route has traffic, pickups in this area take longer). Together, they cover both the predictable physics and the unpredictable context. I also computed an "improvement ratio" — how much the current delivery context deviates from expectation — which gives the model a normalized signal about whether conditions are better or worse than average.

For rider profiling, I aggregated historical performance per rider: median speed, median duration, distance variability, and a composite quality score. Experienced riders with consistent speeds produce more predictable deliveries. For location intelligence, I applied K-Means clustering on all pickup and destination coordinates, creating geographic zone features that capture neighborhood-level patterns without requiring explicit area knowledge.

Feature selection used Weight of Evidence (WOE) analysis on a binarized target (above/below a duration threshold) to identify which features carry predictive signal versus noise. The final model combined H2O AutoML with XGBoost stacking — letting H2O explore the model space broadly while XGBoost provided a tuned gradient-boosted baseline.

Key Decisions

Key Takeaway

The core lesson from this project is that domain knowledge and machine learning are not competing approaches — they're complementary layers. The physics formula sets the baseline. The ML model corrects the baseline using context. Together, they outperform either alone by a significant margin. Whenever a domain formula exists (however imperfect), use it as a feature rather than discarding it. Let the model learn what the formula gets wrong rather than rediscovering what it gets right.