About the Creator

At the intersection of sales strategy and data science

Stephen Murphy
Sales & Marketing Data Science Machine Learning MBA Chicago Booth
Connect on LinkedIn GitHub

My Journey

With over a decade of experience at the intersection of sales, marketing, and advanced data analytics, I've always been driven by a simple question: How can we do this better?

  • Data science enthusiast: I compete in ML competitions and once won an image recognition competition just for fun.
  • Competitive learner: I'm constantly acquiring new skills and solving new problems—this package was developed in just two days from scratch.
  • Practical innovator: I thrive at the intersection of business and technical domains, giving me unique perspective on what practical solutions should look like.
My goal has always been to bridge the gap between sophisticated data science and practical business needs. The best models aren't necessarily the most complex—they're the ones that solve real problems.

How It Started: The Prophet Story

Before diving into Murphet, it's important to understand Prophet. Developed by Facebook's (now Meta's) Core Data Science team and released in 2017, Prophet quickly became one of the most popular time series forecasting tools available.

What Prophet Does Well:

  • Handles missing data and outliers with remarkable robustness
  • Automatically detects seasonal patterns (daily, weekly, yearly)
  • Requires almost no feature engineering to get a solid baseline
  • Doesn't require time series to be stationary (unlike ARIMA models)
  • Has intuitive parameters for business users without statistical expertise

Prophet uses an additive model with three main components: trend, seasonality, and holiday effects. Under the hood, it employs Stan, a probabilistic programming language, for fast model fitting.

Prophet is an exceptional tool, but as I worked with it for specific business cases involving rates and proportions, I began to notice opportunities for improvement.

Why I Created Murphet

Working with bounded metrics like churn rates, conversion percentages, and occupancy rates, I kept encountering four key limitations:

⚠️ Boundary violations

Prophet's Gaussian model allows predictions outside logical bounds (>100% or <0%), requiring manual capping

⚠️ Unrealistic uncertainty

Prediction intervals don't naturally narrow near boundaries

⚠️ Artificial kinks

Hard changepoints create unrealistic breaks in forecasts

⚠️ Missing pattern persistence

The model doesn't capture autocorrelation in residuals

That's when I built Murphet in just two days. The key innovations:

  • Beta likelihood model instead of Gaussian, ensuring predictions naturally stay within (0,1) range
  • Smooth logistic transitions between trend segments
  • Latent AR(1) component to capture pattern persistence
  • Heteroscedastic precision that adapts to data levels

"I'm not claiming Murphet is better than Prophet for everything — it's just better for specific use cases. And that's the beauty of open source: building specialized tools that solve specific problems exceptionally well."

Open Source & Looking Forward

I've open-sourced Murphet because I believe in community validation and improvement. The benchmarks showing 42-56% error reduction for bounded metrics are impressive, but I want to see if these results hold across diverse datasets.

Ideal Use Cases for Murphet:

  • Churn rates and customer retention
  • Conversion rates for e-commerce
  • Hotel and rental occupancy forecasting
  • Market share predictions
  • Any metric that must remain between 0% and 100%

This project represents what I love most about the intersection of business and data science: taking established tools and refining them for specific use cases where they can deliver exceptional value.

If you're interested in contributing or have suggestions, please reach out via GitHub or connect with me on LinkedIn. Let's make forecasting bounded metrics more accurate together.

Get Started with Murphet Contribute on GitHub