The more stores or outlets you have in a network, the more sophisticated the technology becomes when making sales predictions. The more accurate you can be in assessing a new or proposed site, the more comfortable you will be with your franchisee appointment, and the less likely you are to encounter a legal problem.
The larger networks around can move beyond the normal statistical approach of regression modelling, and into the newer approaches such as Neural Networks.
What is a Neural Network you may ask?
Well, in the world of science fiction, Arnold Schwarzenegger's character had a Neural Network in his head when he was in Terminator 2. He could not let this futuristic technology fall into the hands of the people of the present, so had to destroy himself at the end of the film!
We have obviously moved forward since Terminator 2, as Neural Networks are with us today, and definitely not by way of Arnold delivering them to us.
The Neural Network (as the name suggests) is loosely modelled after the human brain. They "learn" from a process where if predictions are improved it gives itself a pat on the back, and if it is not improved, it kicks itself in the backside.
The Neural Network strives to come up with better correlations between the actual sales and the predictions they create. Technically, it is made up of interconnected processors that by changing their connection (called training) learn the solution to a problem. THe end result is an enhanced ability to analyse a large amount of information far better.
In large networks (100+ stores), Neural Networks usually can create a predictive model far more superior and quicker than those derived from regular statistical analysis. There are a large number of factors which affect sales and these factors are often interrelated. Neural Networks are able to recognise these subtle interrelationships between factors, and thus add to the accuracy and precision of predictions.
The short turn around time in Neural Network modelling is because the analyst does not need to try different permutations and combinations to arrive at a good model — the Neural Network algorithm does this and comes up with the best results. All one has to be careful about is the kind of factors fed into the Neural Network, as the old modelling principle of junk-in-junk-out applies here too. Once the variables having effect on sales are determined, the Neural Network can also determine their importance.
There are many funny stories about the kinds of trends Neural Netwrks are capable of recognising and how they can be put to good use. A large supermarket chain in the US used Neural Networks to analyse customer buying habits. The analysis revealed beer and diapers sold well together. It was theorised the reason for this was fathers were stopping off at the supermarket to buy diapers, and since they could no longer go down to the pub as often, would buy beer as well. As a result of this finding, the beers were placed next to the diapers, resulting in increased sales for both.
In another case, stockbrokers on Wall St used Neural Networks to see the relationship between the trends in sales in the last 15 minutes from the closing bell. The intelligence gained enabled the stockbrokers to buy / sell in the last five minutes before closing with a degree of confidence in the final close prices for the day.
Neural Networks need to be tested before they are put to use, this is done to make sure the model will operate as intended. One way to do this is to set aside 25 percent of the stores and build the model on the other 75 percent. This method does not use all available data. For this reason, the model is trained and tested on the same data in turns, i.e. a different sample is left out each and a model made on the rest, till all available data is used. These models are then averaged or combined to obtain a final model.
Neural Networks have been used for around 13 years in the oil industry as a tool to assist in sales predictions for service stations. With our years of experience in sales prediction modelling, there is now the ability to work through the complexity of Neural Networks and present the findings to levels of management in an easy to understand manner. Experiences with Neural Networks has also shown us there is an expectation of around 5-10% improvement in the predictive accuracy of models, compared to regression modelling on the same data set.
So for large networks who want a competitive edge, it's perhaps time to take the next technological leap with a Neural Network. This would help obtain better quality information and analysis from the available data, which in turn would lead to more informed and effective decisions.
By Peter Buckingham and Anubhav Tewari - Spectrum Analysis Australia
22-Oct-2008