Increasing NYC Taxi Drivers Earnings

Problem Description

Opportunity

Taxi drivers could increase their earnings by changing their strategy.

Questions to solve

  • How much can a taxi driver increase their monthly earnings just by skipping trips under defined conditions?
  • How much can a taxi driver increase their monthly earnings just by changing their initial zone and time?

Business success criteria

Develop a strategy to increase NYC taxi drivers’ monthly earnings by 20%.

Project scope

This project will be limited to Juno, Uber, Via and Lyft taxi drivers who work in New York City in trips that take place between any zone of Manhattan, Brooklyn or Queens (the more active ones).

Results Highlight

πŸ€– Modeling for Decision Support

To ensure the model supports real‑world driver decisions, we framed the problem as a sequential decision task rather than a pure prediction exercise. The goal is to maximize net hourly earnings, not just classification accuracy.

  • From probabilistic predictions to profit‑driven thresholds: We selected the Brier score as the primary evaluation metric because it measures both discrimination and calibrationβ€”essential when probabilities inform acceptance decisions. To translate predicted probabilities into actionable trip recommendations, we performed a threshold optimization that accounts for asymmetric misclassification costs: rejecting a lucrative trip (false negative) is far costlier than accepting a mediocre one (false positive). By simulating driver earnings across 5‑fold cross‑validation, we identified an optimal decision threshold of 0.22β€”well below the default 0.5. This lower threshold reflects the fact that drivers should be slightly more willing to accept trips, as the opportunity cost of waiting for a β€œperfect” fare outweighs the risk of a low‑value ride.
  • Robustness and generalisation: The final XGBoost model was evaluated on a held‑out test set, achieving a Brier score of 0.146 (matching cross‑validation) and delivering a net benefit of $0.35 per trip compared to a β€œtake‑all” baseline. This represents a 51.5% improvement in expected cost, confirming that the cost‑benefit trade‑off remains stable on unseen data.

🎯 Sequential Decision Framework & Business Logic

The core of this project is not a simple classification task; it is a Sequential Decision Analytics problem. I transformed raw observational taxi data into a decision-making tool by addressing two major challenges:

  • The Target Variable Dilemma: The original dataset lacked a β€œground truth” for whether a trip was optimal. I engineered the target variable take_current_trip by calculating the Opportunity Cost of each trip. This involved simulating the potential earnings of waiting for a high-value fare versus accepting the immediate request, creating a decision-centric label from scratch.
  • The Baseline Paradox: I established a β€œTake-All Policy” (accepting every trip) as the baseline. While the ML model shows strong predictive performance (AUC and Brier Score), I have framed the project to acknowledge that predictive accuracy does not automatically equate to policy superiority. The model is designed to optimize a threshold that maximizes net hourly earnings, not just β€œhits and misses.”

πŸ’Ύ Engineering for Big Data & Software Reliability (Out-of-Core Processing)

I architected a robust, production-grade pipeline designed to handle datasets exceeding available RAM while maintaining strict software engineering standards:

  • Custom Tidymodels Extensions: To integrate geospatial features seamlessly into the machine learning pipeline, I developed a custom recipes step. This allows for the automated preprocessing of coordinates and spatial joins within a unified workflow, ensuring that feature engineering is consistent during both training and inference.

  • Production-Grade R Development: To ensure reliability, the project is structured as a formal R package, moving beyond simple scripts to a maintainable codebase.

    • Unit Testing: I implemented a comprehensive suite of tests using testthat to validate custom logic, specifically for the simulation functions and the custom recipes steps.

    • Rigorous Documentation: All core functions are fully documented using roxygen2, providing clear API definitions, parameter requirements, and usage examples.

  • Hybrid Analytical Engine: I utilized DuckDB as an out-of-core engine to perform heavy aggregations and joins directly on disk. Once filtered, I leveraged data.table in R for ultra-efficient in-memory manipulation, combining SQL’s disk performance with R’s functional programming power.

  • Reproducible Environments: The entire stack is managed via Nix and Docker, ensuring the environmentβ€”including complex geospatial system dependenciesβ€”is 100% reproducible across any machine.

πŸ—ΊοΈ High-Dimensional Feature Engineering & Spatial Intelligence

  • Conceptual Clustering (NLP): To navigate the β€œCurse of Dimensionality” presented by 160,000+ US Census variables, I didn’t use arbitrary selection. I applied NLP (Jaccard Distance) and Edge-Betweenness clustering to group variables into conceptual themes (e.g., β€œCommuting Habits,” β€œWealth Distribution”), allowing for a data-driven prioritization of features.
  • Geospatial Intersections: Integrated OpenStreetMap (OSM) data by performing complex spatial intersections. I mapped road lengths and amenity densities (restaurants, transit hubs) to specific taxi zones to capture the geographic β€œDNA” of NYC.

Methodology

To find the optimal solution for those questions, we will follow the methodology proposed by Warren B. Powell (2022) in Sequential Decision Analytics and Modeling: Modeling with Python and combine it with the Cross-Industry Standard Process for Data Mining (CRISP-DM) to define a machine learning model that will power the sequential decision to optimize.

Sequential Decision Modeling Framework

CRISP-DM Process Diagram

Following the steps of both methodologies, we have organized the articles created in this portfolio website:

Sequential Decision Analytics CRISP-DM Article Name
Core Elements of the Problem Business Understanding 1. Business Understanding Overview
Data Understanding 2. Data Collection Process
From Defining Mathematical Model to Evaluating Policy Business Understanding 3. Defining Base Line
Data Understanding 4. Data Sampling
5. Initial Exploration
Data Preparation 6. Expanding Geospatial Information
7. Expanding Transportation and Socioeconomic Patterns
Modeling and Evaluation 8. Training and Selecting Model To Implement
From Defining Mathematical Model to Evaluating Policy 9. Implementing ML Model To Sequential Problem (Pending)
Deployment 10. Wrap Decision Model into REST API (Pending)
11. Serving Model by a Shiny Web App (Pending)

Data to Use

In this project, we will use a subset of the data available in the TLC Trip Record Data from 2022 to 2023 for High Volume For-Hire Vehicle β€” which covers the Juno, Uber, Via and Lyft trips within our project scope β€” with the columns described in its data dictionary.

Disclaimer

This project was completed under strong assumptions given that the data used in the analysis does not provide any unique identifier for taxi drivers, which limits the realism of some results.

Additionally, this project aims to increase taxi driver earnings at the individual level. However, if applied extensively, it could also produce the following unintended consequences:

  1. Reduced service quality: Drivers focusing solely on maximizing earnings may avoid less profitable areas or times, potentially leaving some passengers underserved.

  2. Increased congestion: Drivers congregating in high-profit areas could worsen traffic in already busy parts of the city.

This project is intended as a demonstration of data science methodology rather than a prescriptive business recommendation, and these considerations should be carefully weighed before any real-world implementation.

Roadmap & Future Developments

Sequential Decision Analytics (Warren Powell Framework)

The next stage involves moving beyond static evaluation to a dynamic environment. I am integrating the ML model into the framework established by Warren B. Powell (2022) in Sequential Decision Analytics and Modeling.

  • Policy Function Approximation (PFA): Using the ML model’s calibrated probabilities to define a decision policy.
  • Four Elements Modeling: Formally defining the State (driver location, time), Decision (Accept/Reject), Exogenous Information (new trip requests), and the Objective Function (Maximizing total daily revenue).

Deployment & Accessibility

To transition this from a local research project to a production-grade tool, I will implement:

  • REST API: A R/Plumber API to serve real-time trip recommendations based on the trained XGBoost model.
  • Shiny Dashboard: An interactive web application built in Shiny to visualize the driver’s predicted earnings, optimal decision thresholds, and spatial demand heatmaps in real-time.

Try the app in your browser: NYC Taxi Zone Selector on Hugging Face Spaces

Source code: https://github.com/AngelFelizR/nyc-taxi-zone-selector

Project Structure and Tooling

Reproducibility and long-term maintainability were core priorities from the start, which shaped every tooling decision in this project. The following tools were used to achieve this:

  1. We use git to manage changes in the code and provide an interface to share the project on GitHub.
  2. Docker and Nix are used to build a reproducible dev-container based on default.nix. The container can be connected via SSH using a public and private key pair as defined in setup.sh, and the .envrc sets the Nix environment to use in the Positron console.
  3. For modeling, we used the tidymodels framework to ensure we are following good modeling practices.
  4. Since the project follows the basic structure of an R package, we were able to document and create unit tests for custom functions using testthat, roxygen2 and devtools. This was especially important to ensure that the simulation function and the custom step function (which extends the recipes package) work correctly.
  5. The project also follows the structure of a Quarto project and renders all articles into the docs folder, giving us full control over the format used to present each article. Results are hosted on GitHub Pages, so they can be shared at no cost.
  6. The .Rprofile overrides install.packages, update.packages and remove.packages to make clear that R packages must be defined in default.nix to ensure reproducibility.
  7. To manage data larger than RAM, we use duckdb and keep large files in a separate folder named NycTaxiBigFiles under the same parent directory as this repo.
  8. To cache results generated during the investigation process, we use .qs2 files and track them with pins, stored under the folder NycTaxiPins in the same parent directory as this repo.
  9. We use the air extension to ensure consistent code formatting across the project.

The result is a hybrid structure that combines an R package (with documented functions and unit tests) and a Quarto website (with rendered articles and hosted results), which was one of the most challenging aspects of the project to set up correctly:

tree -L 3
.
β”œβ”€β”€ about.qmd
β”œβ”€β”€ air.toml
β”œβ”€β”€ default.nix
β”œβ”€β”€ DESCRIPTION
β”œβ”€β”€ docker-compose.yml
β”œβ”€β”€ Dockerfile
β”œβ”€β”€ docs
β”‚   β”œβ”€β”€ about.html
β”‚   β”œβ”€β”€ figures
β”‚   β”‚   β”œβ”€β”€ CRISP-DM_Process_Diagram.png
β”‚   β”‚   β”œβ”€β”€ htop_parallel_process.png
β”‚   β”‚   β”œβ”€β”€ logo-generated.jpeg
β”‚   β”‚   β”œβ”€β”€ model_benefit_curve.png
β”‚   β”‚   β”œβ”€β”€ model-benefit.jpg
β”‚   β”‚   β”œβ”€β”€ nyc-taxi-navbar-logo.png
β”‚   β”‚   β”œβ”€β”€ nyc-taxi-navbar-logo.xcf
β”‚   β”‚   └── Sequential-Decision-Modeling-Framework.png
β”‚   β”œβ”€β”€ index.html
β”‚   β”œβ”€β”€ investigation-phases
β”‚   β”‚   β”œβ”€β”€ 01-business-understanding.html
β”‚   β”‚   β”œβ”€β”€ 02-data-collection-process.html
β”‚   β”‚   β”œβ”€β”€ 03-base-line_files
β”‚   β”‚   β”œβ”€β”€ 03-base-line.html
β”‚   β”‚   β”œβ”€β”€ 04-data-sampling.html
β”‚   β”‚   β”œβ”€β”€ 05-initial-exploration_files
β”‚   β”‚   β”œβ”€β”€ 05-initial-exploration.html
β”‚   β”‚   β”œβ”€β”€ 06-expanding-geospatial-data_files
β”‚   β”‚   β”œβ”€β”€ 06-expanding-geospatial-data.html
β”‚   β”‚   β”œβ”€β”€ 07-expanding-transportation-socioeconomic_files
β”‚   β”‚   β”œβ”€β”€ 07-expanding-transportation-socioeconomic.html
β”‚   β”‚   β”œβ”€β”€ 08-model-selection_files
β”‚   β”‚   └── 08-model-selection.html
β”‚   β”œβ”€β”€ man
β”‚   β”‚   └── figures
β”‚   β”œβ”€β”€ search.json
β”‚   └── site_libs
β”‚       β”œβ”€β”€ bootstrap
β”‚       β”œβ”€β”€ clipboard
β”‚       β”œβ”€β”€ DiagrammeR-styles-0.2
β”‚       β”œβ”€β”€ ggiraphjs-0.9.2
β”‚       β”œβ”€β”€ girafe-binding-0.9.2
β”‚       β”œβ”€β”€ grViz-binding-1.0.11
β”‚       β”œβ”€β”€ htmltools-fill-0.5.8.1
β”‚       β”œβ”€β”€ htmlwidgets-1.6.4
β”‚       β”œβ”€β”€ jquery-3.6.0
β”‚       β”œβ”€β”€ leaflet-1.3.1
β”‚       β”œβ”€β”€ leaflet-binding-2.2.3
β”‚       β”œβ”€β”€ leafletfix-1.0.0
β”‚       β”œβ”€β”€ Leaflet.glify-3.2.0
β”‚       β”œβ”€β”€ leaflet-providers-2.0.0
β”‚       β”œβ”€β”€ leaflet-providers-plugin-2.2.3
β”‚       β”œβ”€β”€ proj4-2.6.2
β”‚       β”œβ”€β”€ Proj4Leaflet-1.0.1
β”‚       β”œβ”€β”€ quarto-html
β”‚       β”œβ”€β”€ quarto-nav
β”‚       β”œβ”€β”€ quarto-search
β”‚       β”œβ”€β”€ rstudio_leaflet-1.3.1
β”‚       └── viz-1.8.2
β”œβ”€β”€ figures
β”‚   β”œβ”€β”€ CRISP-DM_Process_Diagram.png
β”‚   β”œβ”€β”€ htop_parallel_process.png
β”‚   β”œβ”€β”€ logo-generated.jpeg
β”‚   β”œβ”€β”€ model_benefit_curve.png
β”‚   β”œβ”€β”€ model-benefit.jpg
β”‚   β”œβ”€β”€ nyc-taxi-navbar-logo.png
β”‚   β”œβ”€β”€ nyc-taxi-navbar-logo.xcf
β”‚   └── Sequential-Decision-Modeling-Framework.png
β”œβ”€β”€ generate_env.R
β”œβ”€β”€ index.qmd
β”œβ”€β”€ investigation-phases
β”‚   β”œβ”€β”€ 01-business-understanding.qmd
β”‚   β”œβ”€β”€ 02-data-collection-process.qmd
β”‚   β”œβ”€β”€ 03-base-line.qmd
β”‚   β”œβ”€β”€ 04-data-sampling.qmd
β”‚   β”œβ”€β”€ 05-initial-exploration.qmd
β”‚   β”œβ”€β”€ 06-expanding-geospatial-data.qmd
β”‚   β”œβ”€β”€ 07-expanding-transportation-socioeconomic.qmd
β”‚   └── 08-model-selection.qmd
β”œβ”€β”€ man
β”‚   β”œβ”€β”€ add_performance_variables.Rd
β”‚   β”œβ”€β”€ add_pred_class.Rd
β”‚   β”œβ”€β”€ add_take_current_trip.Rd
β”‚   β”œβ”€β”€ calculate_costs.Rd
β”‚   β”œβ”€β”€ collect_predictions_best_config.Rd
β”‚   β”œβ”€β”€ compare_model_predictions.Rd
β”‚   β”œβ”€β”€ figures
β”‚   β”‚   β”œβ”€β”€ logo.hex
β”‚   β”‚   β”œβ”€β”€ logo-image.png
β”‚   β”‚   β”œβ”€β”€ logo.png
β”‚   β”‚   └── Logo-source.txt
β”‚   β”œβ”€β”€ NycTaxi-package.Rd
β”‚   β”œβ”€β”€ plot_bar.Rd
β”‚   β”œβ”€β”€ plot_box.Rd
β”‚   β”œβ”€β”€ plot_heap_map.Rd
β”‚   β”œβ”€β”€ plot_num_distribution.Rd
β”‚   β”œβ”€β”€ required_pkgs.step_join_geospatial_features.Rd
β”‚   β”œβ”€β”€ simulate_trips.Rd
β”‚   └── step_join_geospatial_features.Rd
β”œβ”€β”€ multicore-scripts
β”‚   β”œβ”€β”€ 01-fine-tune-future-process.R
β”‚   β”œβ”€β”€ 02-add-target.R
β”‚   β”œβ”€β”€ 02-run_add_target.sh
β”‚   β”œβ”€β”€ 03a-tuning-simple-models.R
β”‚   β”œβ”€β”€ 03b-tuning-dimreduction-models.R
β”‚   └── 03c-tuning-tree-models.R
β”œβ”€β”€ NAMESPACE
β”œβ”€β”€ params.yml
β”œβ”€β”€ _quarto.yml
β”œβ”€β”€ R
β”‚   β”œβ”€β”€ add_take_current_trip.R
β”‚   β”œβ”€β”€ calculate_costs.R
β”‚   β”œβ”€β”€ compare_model_predictions.R
β”‚   β”œβ”€β”€ NycTaxi-package.R
β”‚   β”œβ”€β”€ plot_bar.R
β”‚   β”œβ”€β”€ plot_box.R
β”‚   β”œβ”€β”€ plot_heap_map.R
β”‚   β”œβ”€β”€ plot_num_distribution.R
β”‚   β”œβ”€β”€ simulate_trips.R
β”‚   β”œβ”€β”€ step_join_geospatial_features.R
β”‚   └── utils.R
β”œβ”€β”€ README.md
β”œβ”€β”€ result -> /nix/store/63jxvg9zwnwab3jmv74pdsp6pmr2hbww-nix-shell
β”œβ”€β”€ setup.sh
└── tests
    β”œβ”€β”€ testthat
    β”‚   β”œβ”€β”€ fixtures
    β”‚   β”œβ”€β”€ test-add_take_current_trip.R
    β”‚   β”œβ”€β”€ test-calculate_costs.R
    β”‚   β”œβ”€β”€ test-plot_box.R
    β”‚   β”œβ”€β”€ test-simulate_trips.R
    β”‚   └── test-step_join_geospatial_features.R
    └── testthat.R

43 directories, 90 files

Defining Development Environment

To reproduce the results of this project, follow these steps to set up the same environment using Docker and Nix.

1. Install Docker and Docker Compose

You need Docker and Docker Compose. Choose the appropriate installation method for your operating system:

For Debian 13 (as an example), run the following as root:

apt update
apt install -y apt-transport-https ca-certificates curl gnupg2 software-properties-common
curl -fsSL https://download.docker.com/linux/debian/gpg | apt-key add -
add-apt-repository "deb [arch=amd64] https://download.docker.com/linux/debian trixie stable"
apt update
apt install -y docker-ce docker-compose-plugin
systemctl enable docker && systemctl start docker
usermod -aG docker <YOUR-USER>
su - <YOUR-USER>

Note: Replace <YOUR-USER> with your actual username.

2. Clone the Repository and Prepare Directories

Navigate to the parent directory where you want to store the project and the data folders. Then run:

cd <parent-dir-path>
mkdir NycTaxiBigFiles
mkdir NycTaxiPins
git clone https://github.com/AngelFelizR/NycTaxi

Your directory structure should look like:

<parent-dir-path>/
β”œβ”€β”€ NycTaxi/               # cloned repository
β”œβ”€β”€ NycTaxiBigFiles/       # large data files (mounted into container)
└── NycTaxiPins/           # pin board storage (mounted into container)

3. Run the Setup Script

The repository includes a setup.sh script that automates all remaining steps: pulling the image, starting the container, and configuring SSH key-based authentication using your existing ~/.ssh/id_rsa.pub.

From inside the NycTaxi folder, run:

cd NycTaxi
chmod +x setup.sh
./setup.sh

The script will:

  • Pull the pre-built image angelfelizr/nyc-taxi:4.5.2 from Docker Hub.
  • Start the container in detached mode, mapping port 2222 for SSH and mounting the three directories under /root/.
  • Register your public key (~/.ssh/id_rsa.pub) inside the container so you can connect without a password.
#!/bin/bash
docker compose pull
docker compose up -d
docker compose cp ~/.ssh/id_rsa.pub nyc-taxi:/root/.ssh/authorized_keys
docker compose exec nyc-taxi chown root:root /root/.ssh/authorized_keys
docker compose exec nyc-taxi chmod 600 /root/.ssh/authorized_keys
echo "Listo! Conectate con: ssh NycTaxi"

You can verify the container is running with docker compose ps.

4. Configure SSH

Add the following to your ~/.ssh/config so you can connect with a simple alias:

Host NycTaxi
    HostName 127.0.0.1
    User root
    Port 2222
    IdentityFile ~/.ssh/id_rsa

Then connect with:

ssh NycTaxi

5. Using Positron (or VS Code) with direnv

Since direnv is configured via the .envrc file in the repository, you can use Positron with the SSH remote development feature to work directly inside the container.

  1. In Positron, select β€œConnect to Host…” (or use the Remote Explorer).
  2. Enter root@localhost:2222 and authenticate using your SSH key (configured in Step 3).
  3. Once connected, open the folder /root/NycTaxi.
  4. Install the direnv extension by mkhl from the Open VSX Registry. This extension automatically activates direnv when you open a folder containing an .envrc file.

After the extension loads, you should see a notification confirming that direnv is active. At that point, any terminal you open inside Positron will have the Nix environment loaded automatically.

To make the R interactive console use the Nix environment instead of the system default, open the Positron command palette and switch the active R interpreter to the one provided by the Nix shell. Once selected, the console will have access to all the R packages defined in default.nix.

6. Remote Pin Board (Optional)

If you need to use the shared pin board, create a cache directory on your host (outside the container) and then, inside R, set up the board as follows:

# On your host (in <parent-dir-path>)
mkdir NycTaxiBoardCache

In your R session (inside the Nix shell), use:

BoardRemote <- board_url(
  "https://raw.githubusercontent.com/AngelFelizR/NycTaxiPins/refs/heads/main/Board/",
  cache = here::here("../NycTaxiBoardCache")
)

The cache directory is mounted into the container at /root/NycTaxiBoardCache, so pins will be stored on your host and persist between container restarts.