Quantifying Sales Force Performance Using Neural Network Analysis

Assessing individual salesmen's performance is an enormously difficult problem. Outside the legal profession, few people are more persuasive than salesmen, and when it comes to their compensation, no topic is dearer to their heart. "It's totally different in Basingstoke" they cry, and of course it is. Actually, it's different everywhere...

We at the Alexander Partnership believe in managing people by objectives, against controllable performance parameters. We have done a lot of work in the area of retail network modelling, for example, so that clients can assess individual outlet management performance net of external or uncontrollable factors.

The customer demographics in the catchment area of a retail site are clearly outside the control of the outlet manager, (but would be a controllable for the business development manager that selected the site). Through quantitative modelling, the underlying management performance of a retail site can be measured against its peers, with the effects of customer demographics, (and other uncontrollables), taken out. Shops with below average profit performance often possess demonstrably above average management, as a result of an historically P&L-driven culture forcing them to pare costs and exploit every opporunity just to become "average".

When a client asked us to determine the relative performance of the individuals in a direct sales force we jumped at the chance to apply our analytical skills in a similar way.

Defining the problem is the first step in quantitative modelling: we need to decide exactly what we are trying to model first and then list all the influencing factors. One approach would be to model sales value, (per individual), another would be gross margin, but the best would be profit contribution after customer servicing costs. The latter approach takes account of the fact that lots of small orders may have better gross margin but are generally more costly to service ! in invoicing, despatch, credit control and so on, thus enabling measurement of a salesman's contribution to the company's bottom line. Keeping things simple, (always a good idea), we decided to model sales revenue per salesman as a first step, in keeping with traditional client objectives for the sales team.

A number of factors will influence a salesman's performance in his territory, compared to other salesmen in theirs. The territory itself, just as in the retail sites we were accustomed to dealing with, was the first element. Encapsulating the nature of each territory in a series of quantitative measures which might impinge on a salesman's effectiveness prompts the use of total population, average per capita wealth, number of sales points, (in this instance, giftware outlets), population density, consumer expenditure, number of visitors per annum and visitor spending estimates.

Gathering data is often a major hurdle, sometimes exacerbated by a desire to be overly accurate. One lesson we have learned is that if you gather rough but readily available data first, initial modeling will show whether some of the data is important enough to spend time refining. Sometimes factors which are brainstormed as being potentially important influencers prove to be adding little to overall understanding, and can be dropped.

There are several ways in which the data can be related to produce a model of salesmen's performance. One is stepwise regression, whereby the factor most highly correlated with the "output" being modeled is regressed against it first, and the best of each of the other factors is regressed against the successive residuals, in turn. Statistical packages from SPSS or SAS make this relatively straightforward.

The approach we most commonly use, because it's unbiased and gives robust solutions, is to use a neural net inference engine. Neural nets have the ability to deliver the goods with relatively small data sets - and sales forces don't generally run to very large numbers. The keys to success with neural nets are the learning and stopping algorithms, and the software we have helped to develop is particularly effective here.

Having built the best model possible with the available data we are able to assess the likely underlying ability of individuals in the sales team. The model is a formula for predicting what an averagely effective salesman should achieve in a particular territory. If we feed into the model the parameters for any territory and then compare the output with what the relevant sales person is actually achieving we can quantify how much better - or worse - than average she or he is doing.

There is still room for discussion, though; it may be that a territory has only recently been established, or was just taken over from an ineffective representative, and so a period of above average missionary work will be required before "normality" is reached. But the whole dialogue is infinitely better informed as a result of the analysis.

We view this as an important testing phase for any modeling work. New factors - such as the history of a territory's sales efforts - are unearthed, and the modelling aproach can be refined to take them into account, so that the final model has buy-in from both sides of the table.

In this instance we conducted a separate time series analysis of territory performance to estimate the average time taken by a new salesman to reach "normality"; the number of instances of staff changes were not large, but the quantitative evidence again enriched the discussion between sales manager and certain members of his team.

As a precursor to a more extensive "Value Chain Analysis" of the client company we repeated the modeling process with the client's average customer account size included as an input, alongside the other territory sales influencing factors described earlier. This really got things humming, since it suggested that some of the "best" salesmen were apparently merely "coasting" with relatively few big accounts in territories which were underdeveloped. Cherry picking makes good sense, as long as you don't just pick in the morning and take afternoons off...