Not So Slippery Slope

Compass Minerals (NYSE: CMP) produces rock salt for roadway de-icing. CMP represents a potential long-term equity investment, but the large fluctuations in sales, the result of changes in winter weather, increase investment risk by masking potential changes to the underlying business. This project reduces that investment risk, by estimating changes in sales net of the influence of winter weather fluctuations.

Problem Statement

Compass Minerals (NYSE: CMP) produces rock salt for roadway de-icing. The fundamentals of the business are attractive. CMP produces a necessity good with a competitive moat. Due to the very high weight-to-value ratio of de-icing salt (a metric ton of salt sells for approximately $70 in the snow belt region) CMP’s sales in an area are largely determined by distance from their salt mines. The company has a monopoly or oligopoly in all of its markets. The problem is that the company regularly experiences huge fluctuations in sales and cash flow on a year over year basis as the result of fluctuations in winter weather severity. As an illustration, their forward EBITDA guidance for 2023 is $175mm-275mm, a 40% relative range. The average relative range for company’s this size is 10% (source: marketbeat.com). These fluctuations create investment risk because they make it very difficult to monitor for changes to the underlying business. Given that CMP has a cash-cow business there is a very real risk that management will deteriorate over time. This project solves this problem by providing estimates of company performance, via a web application, net of the effects of winter weather severity. Portfolio managers will be able to use the application to get an estimate of quarterly sales growth/decline that is over and above changes in winter weather severity. Having this product will lower the risk of the investment and allow managers to take larger positions, all else being equal.

Methods Used

• Machine Learning
• Geospatial Analysis
• Data Visualization
• Predictive Modeling
• Regression Analysis: Histogram Gradient Boosting Regressor with poisson loss function

Technologies

• Python
• Geopandas, Shapely, Pyproj
• Pandas
• Scikit-learn
• Matplotlib
• Streamlit

Project Description

There are two main steps in the modeling process:

1. prediction of geocoded quarterly market salt usage based on snow-related conditions and extent of roads
2. prediction of company quarterly sales based on predictions in step 1 weighted by distance from closest company depot.

Data sources:

• SNODAS: Geocoded snow-related variables
  • Snow Data Assimilation System (SNODAS) Data Products at NSIDC https://nsidc.org/data/g02158/versions/1
• BTS NAR: Geospatial Roads Data
  • US DOT BTS North American Roads Data https://data-usdot.opendata.arcgis.com/datasets/usdot::north-american-roads/about
• SALT IOWA: Geocoded salt data by winter storm event
  • Iowa DOT Winter Storm Analysis Historic View Salt usage: https://public-iowadot.opendata.arcgis.com/datasets/IowaDOT::winter-storm-analysis-historic-view/about
• Geocoded CMP Salt Depot Locations (estimated from published reports)
• CMP quarterly road de-icing salt volumes (collected from company financial reports)

ML Model 1: Predict quarterly market salt usage by geography

• Overview:
  • Label: total salt usage (lbs.) per salting event, per geospatial polygon (10 km x 10 km)
  • Features:
    • SNODAS variables the day of and day prior to salting event per geospatial polygon.
    • Extent and type of road lane kms per geospatial polygon
  • Estimator: Histogram Gradient Boosting Regressor with poisson loss function
• Data wrangling and munging:
  • Geospatial overlays: geospatially align SNODAS, salt and road variables.
• Feature extraction/engineering:
  • Optimize size of geospatial grids, aggregations of SNODAS variables (max, min, mean), time extent of SNODAS variables, road extent aggregation, and subset of SNODAS and road variables based on correlation matrices
• Cross-validation and hyperparameter tuning

ML Model 2: Predict quarterly company salt volumes

• Overview:
  • Label: quarterly salt sales
  • Features:
    • Market salt usage estimates per geospatial polygon (predictions using fitted Model 1)
    • Depot distances: distance from closest salt depot, per geospatial polygon
• Estimators: economic gravity model weighting sales by distance from depot; linear regression
• Data wrangling and munging:
  • Geospatial overlays: geospatially align SNODAS, salt estimates, and road variables.
• Hyperparameter tuning: distance weights

Application and visualizations

• Streamlit application: src/app.py
  • Example Bar plot: comparison of estimated quarterly volume based on winter weather vs. actual sales volume
    
    
  • Example Salt color map: estimated salt volumes per geography for selected quarter and prior year quarter
    
    
  • Example solid precipitation color map: solid precipitation per geography for selected quarter and prior year quarter
    
    

Getting Started

For development:

1. Clone this repo (for help see this tutorial).
2. Raw data (not constructed with Makefile) kept here: data/raw/
3. Create and activate venv
4. Install dependencies
   pip install -r requirements.txt
   
5. Makefile: running make from root directory will output list of make targets in order
   

For application end use:

1. Clone this repo (for help see this tutorial).
2. Processed data needed for application kept here:
   data/processed/ models/
3. Create and activate venv
4. Install dependencies
   pip install -r requirements.txt
   
5. Runn application from command line
   streamlit run src/app.py