ETA Accuracy & Route Quality Analysis

ETA Accuracy & Route Quality Analysis is an end-to-end analytics project that evaluates the accuracy of Estimated Time of Arrival (ETA) predictions and route quality using New York City taxi trip data. Accurate ETA predictions are important for user experience, driver planning, and operations. This project trains a neural network model in PyTorch to predict trip duration, computes accuracy metrics across different segments, identifies common failure modes, and outlines an experiment design for potential production deployment.

The repository is structured for clarity and reproducibility with a clear progression: loading raw data, applying a consistent schema and derived columns, training the model, calculating metrics, and producing analysis and visualizations.

The pipeline begins by loading raw NYC taxi trip data. The data is normalized for column naming differences and cleaned by computing trip duration and Haversine distance. Invalid rows with negative durations, missing coordinates, or unrealistic trips are removed. Clean data is stored in PostgreSQL tables. SQL scripts then define useful derived columns such as hour of day, day of week, and distance buckets, along with train/evaluation splits. A feedforward neural network with distance and temporal features is trained in PyTorch and used to generate predictions. Evaluation metrics including Mean Absolute Error (MAE), Median Absolute Error (MedAE), 90th percentile error (P90), and percentages of trips within ±10% or ±20% of the actual duration are computed. Visualizations like error histograms and calibration plots support deeper analysis. Spark support is optional for processing larger datasets.

The model uses features such as Haversine distance, encoded hour and day of week, distance buckets, and normalized pickup/dropoff coordinates. It is trained with early stopping and evaluated on temporally separated splits to avoid leakage. Metrics are segmented by distance, hour, and day so that performance can be understood across different conditions. Automatic column mapping allows the pipeline to accommodate common schema variations.

This design emphasizes reproducibility and analytical clarity rather than real-time serving. PostgreSQL serves as the source of truth for data and supports flexible SQL queries. The PyTorch model is trained offline and used in batch prediction during evaluation. Operating within a single market (NYC) ties results to known urban characteristics but would require adjustments to compare across different cities. The baseline model only uses trip-level and temporal features without route geometry or real-time traffic, leaving room for improvement in more production-oriented systems.

Common failure modes include short trips that show higher relative error due to fixed overhead, peak travel hours with high traffic variability, dense urban routing complexity, and long trips with mixed highway and city conditions. The project proposes an online experiment comparing a control model with a treatment model that includes segment-aware calibration or additional features. Key evaluation metrics would include MedAE and P90, with guardrails based on driver cancel rates, reroutes, and user satisfaction.

Current limitations include the lack of explicit route information, no real-time traffic integration, and a static model that must be retrained periodically. Future improvements could incorporate richer route-based features, uncertainty quantification, continuous learning, and production-scale components such as feature stores, model serving, and monitoring infrastructure.