Decision-making is as much an art as a science.
CFOs and their teams have always worked hard to supply and analyze the data their companies need to be able to make solid, fact-based decisions.
However, Finance departments have long been constrained by basic forecasting techniques.
More often than not, the underlying data collection process is time consuming and error-prone, and the result often falls short of three main requirements:
- Depth: even in a mythical perfectly integrated single-ERP organization, forecasts are drawn from aggregate figures, obscuring the original information on which they are based.
- Scope: supporting data is mainly historical and collected within the organization (changes in sales patterns during previous periods, production costs, etc) whereas exogenous factors (change in consumption behavior, competitor positioning, legal changes, etc.) may impact the organization to a much larger extent.
- Quality: when little data is readily available, finance executives will tend to make assumptions based on their experience, gut-feeling or whatever may make the figures palatable to the audience they are intended to be presented to. Results may also be biased by a lack of data ‘freshness’.
Not only is the underlying data unsatisfactory, but its processing is suboptimal. All these approximate figures end up being copied and pasted from spreadsheet to spreadsheet and undergo many manual transformations. For instance, applying the Excel ‘trend’ function (linear regression) on historical figures helps create forecasts that seem ‘about’ right.
This approach is obviously misleading:
- It’s a case of garbage in, garbage out. Regardless of the quality of the forecasting process, if the data is not detailed, sufficient, relevant and up-to-date, the result will be inadequate.
- The future is not a mere repeat of the past. Making the assumption that “all other things will remain equal” when drawing a trend line is an over-simplification of the reality of the economic environment given that its main characteristic is volatility.
- No lessons are learned from previous errors and therefore there’s no scope for improvement to the next forecasting cycle.
Data Science means that automated and accurate forecasting and financial modelling is today within every CFO’s reach.
Digitization now gives access to more granular and diverse data – be that quantitative or qualitative, structured or unstructured – about present conditions or past situations and their outcomes.
Any data set that may help describe, explain, predict or even determine a company’s positioning can now be stored, updated and processed. This is the case whether that data has been sourced from an internal system or from the world outside (documents exchanged with business partners, information about the industry and macro-economic environment, relevant comments on social networks, weather conditions, etc.).
This 360° view creates an opportunity to discover correlations between the collected data and the figures tracked by finance executives in their modelling activity. But to derive valuable knowledge from data diversity, the trend line methodology is not sufficient.
For the process of discovery to take place, this newly found trove of data needs to be mined with Machine Learning technology.
To put it simply, Machine Learning is the automated search for correlations or patterns within vast amounts of data. Once a statistically significant correlation is identified with a high degree of certainty, it may be applied to new data to predict an outcome.
Let’s take a simple example. Assume you are the CFO of a company that sells goods to other businesses and you want to anticipate your Customer Payment behavior to prevent delays and accelerate your total inbound cash flow.
The traditional way would be to look at your past transactions and payment experiences with every significant customers and infer a probable date of payment for each of them.
But if you take another approach and look closer at your data, you may well find that your customer payment behaviors are not always consistent across time, that your historical view is missing some essential explanatory information about the customer’s behavior that may or may not be specific to their relationship with your company. You end up shooting in the dark.
Wouldn’t your cash-in forecasts be much better if you had also correlated the actual time your customers took to pay you in the past, with detailed information about those transactions?
- Information about each customer such as:
- their payment behavior with other companies
- their ordering frequency
- the date of their first order
- the date of their previous order
- whether there’s a contract between your customer and your company or not
- applicable payment terms
- Information about each transaction such as:
- the amount of the invoice
- the type of invoiced products (direct vs. Indirect, strategic / commodity)
- whether there was a PO or not
- the accuracy of customer business data on the invoice (PO number, analytical code, …)
- whether delivery of goods was acknowledged or not
- what dunning process was applied to generate payment
- Information about the industry & macro-economic situation such as growth or the level of short term interest rates
In theory, you cannot be sure that this model will perform well until :
- You have run a Machine Learning algorithm on your own data, looking for predictive rules that relate each payment behavior to the detailed information of the corresponding transaction.
- You have tested the predictive power of those rules on a set of examples.