BrainTech, LLC: Using Statistics to build Models of Customer Behavior

BrainTech, LLC specializes in data analysis and statistical modeling.  By combining time-related, seasonal, environmental, and psychological factors that may be associated with an event outcome, the models we build help identify and quantify patterns of behavior in specific groups of individuals.  In addition, these models can be used estimate the frequency and likelihood of short-term future events.

We currently work primarily with local businesses. Combining sales trends, seasonality, and behavioral influences with the number of sales made on any given day allows us to build models that both describe customer behavior and estimate short-term future sales. Combined with proper interpretation, our clients often gain insight into the conscious (and unconscious) decision-making processes of their customers, and are better able to identify and quantify factors that influence product demand.   For companies that produce or sell perishable goods, we help with inventory management by using heavily cross-validated models to estimate the number of products that will be sold each day over the upcoming week.

Combining diverse sources of data to build models that include behavioral and market-related influences to sales is the first step. We provide additional value by properly extracting and interpreting the information contained within each model, and then conveying this information to our clients in a manner that is both immediately useful and easy to understand. 

We also offer optional reference material and informative examples to individuals interested in learning more about statistics and statistical modeling techniques.

---

Technical content below. We gratefully appreciate academic and theory-related comments, input, questions, or suggestions for improvement.   The method we use to build short-term predictive models is outlined below:

1. We use the latest brute-force nonlinear multiple regression computer algorithms to perform exploratory data analysis on portions of historic sales data.
2. Our statistician verifies that the generated models are plausible and uses rigorous cross-validation to ensure that the promising models have at least some potential to predict future sales.
3. Since our predictors are correlated (we have knowingly treated non-orthogonal variables as orthogonal), we untangle correlated variables using various methods of confirmatory analysis and raw computing power.
4. After the previous step, we are usually able to isolate the variables that most likely have influenced sales in the past.  Basically, step #3 addresses the frequent occurrence of a variable such as thunderstorms occurring as a negative influence on sales in a portion of the models, while other models include precipitation as a negative influence on sales (but not thunderstorms).  Since thunderstorms and rain are correlated, hypothesis testing must be used to see if (a) thunderstorms or precipitation was a negative influence on sales, or if (b) something more general (like lousy weather) explains decreased sales better than thunderstorms or precipitation alone.
5. We use our custom nonlinear multiple regression program again to find the model that best fits portions of past sales data. This time, only the variables that consistently appeared in the most promising potential models are considered as predictors. At this stage we use rigorous cross-validation to further ensure that we can be confident that the generated model can forecast future sales with sufficient accuracy.
6. Using hypothesis testing, we systematically check each variable we dropped along the way to make sure we have not overlooked a potential predictor that might explain a significant amount of additional variance in sales.
7. If dropped variables are found to explain a significant additional amount of variance in sales, the model is discarded and the process is restarted.
8. If the model passes step #7, we begin to interpret the potentially-complex interactions.  The end goal is to extract important business-related factors that influence sales.
9. The final step involves finding at least ten different ways to convey the essential meaning of the model to our clients.  We also create illustrative examples to clarify important or complex concepts in terms of the customer's specific product and market.

---