Analysts often find themselves lacking the data they want for training a specific model either by being short of records or by these not being rich enough. Exposing some of the common challenges faced when there is not enough data, we will walk through the creative process in training a real-life not-enough-data forecasting demand model where we had to deal with both cases. Using as input just the weekly sales report of 72 restaurants of a renowned fast food franchise we successfully trained a model for choosing the best coordinates to open new restaurants. In the process we faced many common challenges that we addressed in creative and non-orthodox ways that we will be sharing with the audience. Starting with the reduced amount of data that was available for training the model, we complemented the coordinates of the actual points of sales with info from businesses and points of interest that were around them, leveraging on Google Places. Then we translated all that sparse information into more than 300 numeric variables that all together would describe the environment where the restaurants were located based on the kinds of establishments that were around them. We defined our very own formula that we got to name economic concentration index. Finally, we applied some advanced techniques including Ridge Lasso, backward and forward selection, PCA, and Cross Validation to reduce dimensionality and train an surprisingly good linear regression model that would have a residual error fluctuating between 10 and 300 transactions over an average of 2,800 transactions per restaurant per week. With this model we are able to correctly choose 9 out of 10 times the best place to open a new restaurant based only on the given coordinates A and B.
*Training without big data

Enrich data with other sources
*Creating numeric variables from descriptive sparse information
*Modeling with linear regression
*Forecasting demand

Luis Valdeavellano

Talks

Data Con LA 2020