by Nat Evans, Pitney Bowes Business Insight
Most contemporary predictive models, which forecast performance such as sales, customer visits, membership levels, etc., are based on historical data that create “snapshots in time,” using whatever relevant sources were current at the time of analysis. Examples include POS distributions, store and competitive locations, store sales performance and demographic data. But we know operations and the environment changes as soon as a model is completed and put into use. As a result, model accuracy erodes with each passing day as the data inputs into the model or the benchmarks upon which expected performance are based become stale. To be sure, most site selection professionals and researchers attempt to make sure models are as fresh as possible, updating these data elements on a regular and recurring basis. During recent engagements with several long time clients, we have been asked if there was a way to take into consideration dynamic time series data elements to help with forecasting and minimizing risks.
What do we mean by dynamic data?
Many factors may play pivotal roles in retail forecasting and market prioritization. Depending on the level of aggregation, the obvious thought is that a researcher may be able to affect a change in market conditions or individual sales estimates, depending on the application. Indeed, they can significantly sway analyses enough to change even the simplest of decisions, either minimizing risks (if used appropriately) or increasing a company’s vulnerabilities, especially given the current macro-economic climate.
A couple of sources of dynamic data within the context of a static model may include:
• Macro-economic data such as housing starts, CPI (consumer price indices), funds rates, and unemployment percentages either nationally or at varying levels of macro geography – state, county, or CBSA. Such measures provide a look into the health of consumers’ collective behavior, and depending on how the analysis is structured, whether these factors will be leading or lagging indicators of retail growth and consumer spending (PBBI has created an approach-MarketPulse-that incorporates these factors into predictive models).
• Gas prices. Gas price fluctuations on a regional or even local level can create a similar effect that macro-economic variables may produce in models. Obviously, the higher gas prices rise, the less disposable income consumers will have to purchase goods and services, potentially depressing actual local store performance. Distance may become a stronger deterrent to patronage as a result.
If a retailer’s or restaurant’s sales forecast model was created in better times, it may produce a “false positive,” inappropriately triggering a go/no-go decision and costing company valuable resources and capital from other locations that may be more profitable. Just as importantly, if a company is judging a general or district manager on existing location(s) sales performance based on a projection created earlier in the fiscal year, the company may be unduly influencing that leader’s performance rating on factors outside of his or her control.
How can we create more flexible models using dynamic data?
There exists a myriad of ways we can leverage dynamic data through any forecasting or analytical process, more generally. The important point with any data source is to leverage any and all relationships that may prove fruitful through the forecasting process. But, it must be relevant to your research design, have purpose, and be significant enough to warrant using in modeling and analytical review.
In the future, the ability to collect and cleanse data continuously not only from existing, well-documented sources, but also new sources, such as e-commerce and online social/behavioral data, will become more available and increasingly important across any organization. Additionally, whether on-premise or in the “Cloud”, the technology that facilitates a seamless data flow into predictive applications should enable decision-making with the most up-to-date analysis possible.