Traffic Prediction Project with Google Maps & Waze

Build an end-to-end traffic forecasting pipeline using Google Maps, Waze incidents, and time-series models to optimize routes and save travel time.

Cut traffic predictions from noise to action — a hands-on ML project for students

If you feel lost choosing which data sources and models actually move the needle on real-world routing — you’re not alone. Students and early-career practitioners often waste weeks collecting maps, API keys, and messy GPS logs before they can build a working traffic forecasting pipeline. This tutorial compresses that learning curve into a practical, career-ready project: combine Google Maps routing, Waze crowdsourced incidents, and open map layers to forecast congestion and recommend faster routes.

Why this project matters in 2026

By 2026, on-demand mobility, delivery optimization, and shared-mobility systems depend on accurate short-term traffic forecasts. Recent late-2025 updates from major mapping platforms increased the availability of predictive traffic features and accelerated integration with crowdsourced incident feeds — pushing demand for applied skills that connect routing APIs to ML systems. Employers want developers who can combine API data ingestion, time-series feature engineering, and lightweight models that run in production with strict latency and cost limits.

Project overview — what you’ll build

Ingest real-time and historical traffic/travel-time signals from Google Maps Directions (traffic-aware travel time) and Waze incident feeds.
Map-match GPS/routing polylines to road segments (use Google Roads API or OSRM) and build per-segment time series.
Engineered features: lagged speeds, rolling congestion indices, incident flags, weather and calendar features.
Train baseline and advanced models: LightGBM / XGBoost baseline, an LSTM or Temporal Fusion Transformer for sequence modeling, and a simple Graph Neural Network for spatio-temporal correlations.
Evaluate using rolling-window cross-validation and route-level metrics; simulate routing decisions and measure travel-time savings.
Deploy a simple inference service that answers “predict travel time for route R at t+15m” and returns optimized route choices.

Data sources & access (practical)

1) Google Maps Platform

Use the Directions API (departure_time and traffic_model parameters) to get travel-time estimates and route polylines. For precise geometry and snap-to-road, use the Roads API. Note: Google doesn't provide raw per-lane speed sensors — you infer segment-level travel times by combining polylines and travel_time responses from sampled origins/destinations.

2) Waze crowdsourced incidents

Sign up for Waze for Cities (Connected Citizens Program or the Waze Data Feed) to receive incident streams: accidents, jams, hazards and their severity/timestamp. These events are high-signal predictors for short-term congestion spikes.

3) OpenStreetMap (OSM) or OSMnx

Download road geometry, speed limits, and road class. Use OSMnx to extract a network for your city and get attributes (one-way, lanes, highway type). These attributes become stable features for each road segment.

4) Weather and calendar

Use a weather API (OpenWeatherMap, Meteostat) for precipitation, visibility, and temperature. Add calendar flags: weekday/weekend, local holidays, major events. Weather + events are often multiplicative factors on congestion.

5) Optional: probe data and fleets

If you have access to fleet GPS logs, they’re gold for travel-time labeling. Otherwise, simulate probes by sampling travel-time via the Directions API across staggered departure times.

Quick compliance note: always check API Terms of Service before storing, aggregating, or redistributing third-party mapping data. Waze and Google have explicit restrictions on raw data resale.

Data pipeline — step-by-step

Step 1: Define road segments (the spatial unit)

Extract a consistent set of segments using OSMnx or by snapping Directions polyline edges to a canonical network. Each segment should have a unique ID and geometric centroid. Why segments? They give you reusable time-series and reduce dimensionality.

Step 2: Map-matching and aggregation

Map-match each route polyline (from Directions or probes) to segment IDs using the Roads API or OSRM’s map matching. Aggregate travel times and speeds into fixed time buckets (e.g., 5- or 15-minute intervals) per segment.

Step 3: Join incidents and context

For each segment-time bucket, add incident flags from Waze (binary + severity), weather variables, and calendar labels. Also add metadata: road type, speed limit, typical free-flow speed.

Step 4: Labeling

Your target is usually travel time or a normalized congestion index (observed_speed / free_flow_speed). Consider multi-horizon labels: t+5, t+15, t+30 minutes if you want a planner that optimizes short-term decisions.

Feature engineering — what actually predicts congestion

Feature engineering is the part hiring managers prize most. Below are high-ROI features:

Lag features: speed_{t-5}, speed_{t-10}, speed_{t-15}. Short lags capture momentum.
Rolling stats: rolling mean and std of speed over 15–60 minutes.
Incident flags: is_incident, incident_count, avg_incident_severity in last 30m.
Downstream/upstream influence: average speed of adjacent segments (±1–3 hops). Congestion propagates.
Time features: sine/cosine transforms of hour-of-day, day-of-week, and holiday flags.
Road attributes: lanes, max_speed, highway_class, junction_density.
Weather features: precipitation intensity, visibility index.
Event features: stadium events, roadworks — binary flags with lead/lag windows.

Modeling approaches — quick guide

Start simple, then iterate. Below is a progression with practical trade-offs.

Baseline: Gradient-boosted trees

Tools: LightGBM or XGBoost. These handle heterogeneous features, missing values, and are fast to train. Use them as your initial production candidate. They perform strongly on tabular features built from lag/rolling stats and incident flags.

Sequence models: LSTM / Transformer

Tools: PyTorch / TensorFlow. LSTMs are a reasonable next step for segments with long temporal dependencies. The Temporal Fusion Transformer (TFT) (a 2020s advancement) is a strong fit for multi-horizon forecasting and heterogeneous covariates. Expect higher compute cost and more hyperparameters.

Graph-based models for spatial context

Spatio-temporal Graph Neural Networks (DCRNN, Graph WaveNet, STGCN) model propagation of congestion across the network. Use these if adjacent-segment influence is critical. Frameworks: PyTorch Geometric, DGL. Graph models are more complex but can capture wave propagation effects that tree models miss.

Hybrid approach

Combine a per-segment LightGBM baseline with a lightweight graph correction model that adjusts predictions based on neighbors. This often gives the best cost/performance for student projects.

Training & validation — avoid common pitfalls

Temporal cross-validation: use rolling-window CV, not random split. Traffic is autocorrelated and non-stationary.
Evaluation horizons: report metrics for 5m, 15m, 30m horizons separately.
Metrics: MAE / RMSE for travel-time regression; MAPE is useful but be cautious when denominators are small. For incident forecasting, use precision/recall and F1.
Backtesting routing: Evaluate routing decisions by simulating origin-destination queries over historic windows. Compare total travel time for routes selected by predicted travel times vs baseline (Google routing or historical averages).

Edge cases & robustness

Every student project that impresses employers explicitly handles edge cases.

API rate limits & sampling: Google and Waze APIs have quotas. Implement exponential backoff, caching, and downsample probes to stay within limits.
Missing data: impute using rolling medians or use model-based imputation; tree models handle NaNs gracefully, which helps.
Seasonality shifts: holidays, school terms, and pandemic-like disruptions change baselines. Track model drift and set retraining triggers (error delta thresholds).
Label noise: probe data can be biased (fleet vehicles avoid certain roads). Use stratified sampling to reduce bias.
Privacy & legal: anonymize probe data; adhere to Waze/Google ToS on storing location data and redistributing incident feeds.

From model to route optimizer

Once you can predict per-segment travel times for t+H, the next step is to recompute route travel time by summing predicted travel times across segments for each candidate route returned by a routing API.

Query Google Directions for top-K candidate routes for origin/destination without traffic or with current traffic.
Map-match route polylines to segment IDs.
Aggregate predicted segment travel times for your horizon (add buffer for model uncertainty).
Select route with minimum predicted travel time; optionally incorporate risk-aversion (prefer more reliable routes) using predicted variance or ensemble disagreement.

Simulate many O/D pairs and report travel-time savings and reliability gains. Use paired tests to quantify statistical significance.

Practical compute & deployment tips

Use Colab / Kaggle for prototyping and a single GPU for model training. For production-like deployment, use a small cloud VM or serverless functions for inference.
For low-latency inference (sub-second), serve LightGBM models as REST endpoints or use ONNX runtime for fast LGBM inference.
Cache predicted travel times for short TTLs (e.g., 30–60 seconds) to reduce API calls and costs.
Implement explainability hooks: feature importances, SHAP values for route-level decisions. These are great to show on your portfolio.

Evaluation: what success looks like

Beyond MAE, employers look for product-level impact:

Mean travel-time reduction: average minutes saved per route vs baseline.
Reliability improvement: reduction in travel-time variance or late-arrival probability for deliveries.
Model latency and cost: CPU/memory and per-query cost constraints for production.
Resilience: graceful fallbacks when APIs fail (use historical averages).

Advanced strategies and 2026 trends to consider

In 2026 you should be aware of these developments to future-proof your project:

Spatio-temporal foundation models: Larger pre-trained sequence models for traffic forecasting are emerging. Fine-tuning pre-trained temporal models can reduce your training data needs.
Edge inference: With the growth of edge compute in delivery fleets, lightweight on-device prediction (quantized models) is becoming viable for route adjustments en route.
Privacy-first telemetry: More partners are adopting federated analytics for probe data; explore privacy-preserving aggregation if you work with sensitive fleet data.
Hybrid simulation: Using traffic microsimulators (SUMO) to augment rare-event training data (major accidents, rare closures) helps models handle low-frequency, high-impact cases.

Example project checklist and timeline (4 weeks)

Week 1: Get API access, extract OSM network, implement map-matching, and collect one week of pilot data.
Week 2: Build feature pipeline, aggregate segment-level time series, create labels for 5/15/30m horizons.
Week 3: Train baseline LightGBM, validate with rolling CV, and run backtest route simulations.
Week 4: Iterate with a sequence or graph model, implement a simple inference endpoint, and prepare a demo notebook and README for your portfolio.

Common interview / portfolio talking points

When you present this project to recruiters or hiring managers, highlight:

Data engineering decisions: why you chose segment-level aggregation, map-matching choices, and API sampling strategies.
Feature engineering: which features improved short-horizon predictions the most (quantified).
Model selection and trade-offs: explain why LightGBM vs TFT vs GNN for your use case.
Systems thinking: caching, API quotas, fallbacks, and deployment considerations.
Real-world impact: simulated route-time savings and a demo visualizing predicted congestion on a map.

Code and tooling suggestions

Libraries and tools to speed development:

Data & mapping: OSMnx, geopandas, shapely, pyproj
Map matching: OSRM, Google Roads API
APIs: google-maps-services-python, Waze Data SDK (where available)
Modeling: scikit-learn, LightGBM, XGBoost, PyTorch, PyTorch Lightning, PyTorch Geometric
Deployment: FastAPI / Flask, Docker, ONNX Runtime for low-latency scoring
Visualization: kepler.gl, deck.gl, folium for interactive maps

Example pitfalls from student projects (and how to avoid them)

Collecting only a single week of data: not enough to capture weekly patterns — collect at least 4–8 weeks for baseline models.
Treating incidents as independent: include rolling windows; incidents have lead/lag effects.
Using randomized CV: use temporal CV to avoid optimistic performance estimates.
Ignoring API ToS and quotas: design for graceful degradation and local caching.

Final checklist before demo day

Clear README explaining data sources and legal constraints
Simplified demo notebook: input an O/D pair, get predicted travel times and recommended route
Dashboard or map visualization showing predicted vs observed travel times over a test day
Metrics table with MAE/RMSE and route-level travel-time savings

Key takeaways — build this project to stand out

Traffic prediction projects combine engineering judgment with model design. Employers look for evidence you can obtain and clean map data, engineer high-value features (lags, incidents, neighbor influence), and produce robust, low-latency predictions integrated into a routing decision. In 2026, demonstrating knowledge of spatio-temporal modeling, API constraints, and deployment trade-offs separates a résumé from a hireable portfolio piece.

Next steps & call-to-action

Ready to build? Start by scoping a city-sized dataset: get OSM network, request Waze access, and get a Google Maps API key. If you want a guided path, we publish a complete project repo with notebooks, example pipelines, and a deployment template that you can adapt to your city. Join our cohort at skilling.pro for step-by-step mentoring and a portfolio review that shows hiring managers your traffic forecasting system end-to-end.

Take action now: pick an O/D pair in your city, collect 7 days of sampled travel times with Directions API and Waze incidents, and train a LightGBM baseline. Share your demo and metrics — we’ll review it and help you iterate to production readiness.

Build a Traffic Prediction Project Using Google Maps and Waze Data

Cut traffic predictions from noise to action — a hands-on ML project for students

Why this project matters in 2026

Project overview — what you’ll build