Data Science Case Studies

Author

Pavel Logačev

Published

June 30, 2025

1. Time Series Elasticity Analysis

NumPyro-based Bayesian time series model for sales data analysis. A synthetic sales dataset is constructed with latent growth, seasonal patterns, and a random walk component to simulate external influences. Sales are modeled as a Poisson process with price elasticity effects, ensuring realistic demand shifts. A flexible Bayesian model, implemented in numpyro, is used to estimate the underlying components of the sales time series. Inference is performed using MCMC sampling. Despite structural differences, the model is flexible enough to meaningfully separate conceptually distinct components of the sales time series, such as trend, monthly and weekly seasonality, as well as price elasticity of demand. The results confirm that the model successfully reconstructs these components, demonstrating its suitability for the analysis of sales data.

🔗 Project Documentation + Code

2. Distribution-based Change Point Detection

Bayesian changepoint model for detecting structural changes in pricing regimes. Different implementations of a Bayesian changepoint detection model for detecting structural changes in the data-generating process, such as different pricing regimes. Estimation is based on price histograms over time when the pricing of a product undergoes a meaningful change — such as list price changes, promotions, or changes in the availability of discounts.

🔗 Project Documentation + Code

3. Price Distribution Elicitation from an LLM

Two-stage price distribution elicitation pipeline using GhatGPT. Demonstrates a two-stage pipeline that uses large language models (LLMs) to elicit reasonable price distributions for consumer products when competitor pricing data is insufficient or historical sales data is sparse. It extracts structured product attributes from free-text titles and descriptions (via GPT-3.5-turbo), then uses GPT-4o to elicit a retail price range (minimum, typical, maximum) based on those attributes and regional context. The elicited price ranges are used to reverse-engineer log-normal distribution parameters, which are validated against actual prices scraped from Mercado Libre. The result is a calibrated, probabilistic estimate of product pricing grounded in LLM-elicited knowledge.

🔗 Project Documentation + Code

--- title: "Data Science Case Studies" author: name: "Pavel Logačev" email: "pavel@example.com" date: today format: html: toc: true code-fold: true code-tools: true df-print: paged engine: knitr editor: visual --- # ## 1. Time Series Elasticity Analysis **NumPyro-based Bayesian time series model for sales data analysis.** A synthetic sales dataset is constructed with latent growth, seasonal patterns, and a random walk component to simulate external influences. Sales are modeled as a Poisson process with price elasticity effects, ensuring realistic demand shifts. A flexible Bayesian model, implemented in numpyro, is used to estimate the underlying components of the sales time series. Inference is performed using MCMC sampling. Despite structural differences, the model is flexible enough to meaningfully separate conceptually distinct components of the sales time series, such as trend, monthly and weekly seasonality, as well as price elasticity of demand. The results confirm that the model successfully reconstructs these components, demonstrating its suitability for the analysis of sales data. 🔗 [Project Documentation + Code](time_series_analysis_1) ## 2. Distribution-based Change Point Detection **Bayesian changepoint model for detecting structural changes in pricing regimes.** Different implementations of a Bayesian changepoint detection model for detecting structural changes in the data-generating process, such as different pricing regimes. Estimation is based on price histograms over time when the pricing of a product undergoes a meaningful change — such as list price changes, promotions, or changes in the availability of discounts. 🔗 [Project Documentation + Code](pricedist_changepoints) ## 3. Price Distribution Elicitation from an LLM **Two-stage price distribution elicitation pipeline using GhatGPT.** Demonstrates a two-stage pipeline that uses large language models (LLMs) to elicit reasonable price distributions for consumer products when competitor pricing data is insufficient or historical sales data is sparse. It extracts structured product attributes from free-text titles and descriptions (via GPT-3.5-turbo), then uses GPT-4o to elicit a retail price range (minimum, typical, maximum) based on those attributes and regional context. The elicited price ranges are used to reverse-engineer log-normal distribution parameters, which are validated against actual prices scraped from Mercado Libre. The result is a calibrated, probabilistic estimate of product pricing grounded in LLM-elicited knowledge. 🔗 [Project Documentation + Code](pricedist_elicitation)