Back to the work
Data & analytics / Applied machine learning

Logistics Delay Prediction

The best threshold depends on the problem.

Shipment-delay prediction with model comparison and threshold tuning focused on recall and false negatives.

Data & analytics

A model chosen for the cost of missing a delay.

01Explore data
02Engineer features
03Compare models
04Tune threshold
01

The problem

Missing a shipment delay can matter more operationally than reviewing an extra false alert. The project therefore evaluates models using recall, precision, F1, and false negatives rather than accuracy alone.

02

Model comparison

Logistic regression, decision-tree, and random-forest classifiers were compared on the project dataset. Exploratory work examined waiting time, traffic, inventory, and related operational variables. The repository documents feature selection and limitations.

03

The threshold decision

The reported logistic-regression comparison lowers the decision threshold from 0.5 to 0.3. Recall rises from 0.72 to 0.95, while false negatives fall from 38 to 6. The tuned model reports 0.87 F1 and 0.80 precision. These are project-dataset evaluation results, not a production performance claim.

04

What it demonstrates

The useful engineering decision is the tradeoff between missed delays and false alerts. The project translates that choice into operational recommendations and documents the evaluation so the result can be interpreted in context.

KEEP EXPLORINGSPARCS Healthcare SQL Insights