← All projectsIntelliPath / completed

IntelliPath

Course project · Spring 2026

Machine Learning

An open-ended team project: pick a machine-learning problem tied to the course and your own interests, then carry it from proposal to final report.

The assignment

  • Proposal (2–3 pages): problem, motivation, dataset, method, experiment plan, expected outcome and evaluation criteria.
  • A 15-minute presentation with a demo and key findings.
  • Midterm report (4–5 pages): preliminary results and adjustments.
  • Final report (6–8 pages): the full pipeline, results, evaluation and discussion.

The delivery

IntelliPath: freeway speed forecasting on Caltrans PeMS sensor data with Alex Pulikkottil and Matilda Minasian. The results shown are an extended evaluation, where seasonal LightGBM had the lowest overall error at every horizon.

My contribution

I built the PeMS data pipeline, forecasting evaluation, spatial comparisons and visualization.

Prompt: ML Course Project Guide and project assignments (Canvas)

Predicting LA freeway speeds,
up to an hour ahead.

— freeway sensors, LA & Ventura33 in the study

01Forecast the next hour.

Measured speed and forecasts made from earlier readings. Drag through the day or choose a sensor.

Loading…

measuredforecast, 30 min ahead

02Sealed until the end.

Select on May–July data. Freeze the models. Evaluate August–September without further tuning.

TrainJan 2025 – Apr 2026

DevelopMay – Jul 2026

HoldoutAug – Sep 2026

  • Persistence
  • Last week
  • Historical median
  • Seasonal-adjusted
  • Ridge + seasonal
  • LightGBM
  • LightGBM + seasonal
  • LightGBM + seasonal + spatial

 

03Holdout results.

Lowest overall error across 33 sensors. Congestion-specific results were mixed.

LightGBM + seasonalLightGBMRidge + seasonal

2.1 mph15 min ahead
  1. 2.07
  2. 2.15
  3. 2.27
2.6 mph30 min ahead
  1. 2.63
  2. 2.77
  3. 2.88
3.0 mph1 hour ahead
  1. 3.03
  2. 3.24
  3. 3.35

My part

I built the PeMS data pipeline, forecasting evaluation, spatial comparisons and visualization.

Team of three, with Alex Pulikkottil (baseline framework) and Matilda Minasian (bandit prototype).

Shows

Machine learning
Time-series forecastingGradient boostingSeasonal featuresSealed holdout evaluationWeek-block bootstrap
Data
Sensor data quality
Interface
Data visualization

Built with

PythonpandasNumPyscikit-learnLightGBM

Details

Method LightGBM · seasonal features · held-out evaluation

Data Caltrans PeMS · District 7