Business Modelling – Predictive Analytics

Predictive Analytics is a red hot area that touches on Statistics and Data Science, with a touch of human intuition. From a Business Modelling perspective, Predictive Analytics can add more precision and scale to the modelling process compared to simple tools in the Business Model Toolkit like the Business Model Canvas.

Imagine you are the business looking at the graph above. We reach the end point of the blue line and you have to try and predict:

  • Will sales continue to go up? (Green)
  • Will sales flatten off for a period? (Yellow)
  • Will sales start to go down? (Red)

There are a host of factors that could affect this decision, including:

  • Factors leading to the strong growth – new products, entering new markets, economic conditions, etc.
  • Promotions or Marketing Tactics – promotions on products/services, a new marketing campaign, a new strategy by management
  • Seasonality – the growth happened during a strong seasonal period of the year and is expected to change at the end of the season

Both ‘Sales’ and ‘Time’ are left intentionally unmeasured to illustrate the concept of the above. If we knew the amount of sales (either in dollars or units) and the time scales we are seeing (either daily, monthly, quarterly, or yearly), then it would be much easier to understand the context behind the run-up in sales and try to predict what will come next.

Predictive analytics relies on statistical modelling, data science, and other methods to complement our own intuitions. Simply looking at trend graphs and trying to predict where they go next is much like looking at stock charts and trying to predict the market the next short-term movements in the stock market – it is a low probability play.

Seasoned operators and managers will have their own ‘process’ for how they make decisions on forecasting demand, but even small errors in forecasting can have large consequences on any business:

… some 10% of the forecasting improvements may result in up to 30% of the company‚Äôs saving annually, while forecasting improvement costs are simply negligible compared to the total annual cost savings.

Demand Forecasting, Resource Planning, and Procurement Strategy

Trend, Seasonality, Noise

Let’s imagine that the above graph was actually taken over a 4 month period. This helps to sharpen the context and look for signals/patterns within the data that could shape our view of how to predict what comes next. Generally, with this type of data, we are looking for three key elements:

  • Trend – is the trend increasing or decreasing?
  • Seasonality – are there seasonal factors that cause repeating factors in the data?
  • Noise – is there any random variation or outlier data points in the series?

In the case we see above, we can only see a short-term picture of the data (4 months), meaning that all we can really see is the trend (increasing) – we don’t know if this is indeed a seasonal pattern or if any of the above data points are noise.

One of the key first principles of predictive analytics is that there needs to be enough data (at least 12 months, ideally 24 months) to be able to accurately build a trend forecast. 12 months data may be enough to start to see some potential noise (large spikes up or down), which may be characteristic of a startup or new company that is just entering the market. 24 months would help us to start to see some seasonality emerge.

Predictive Analytics – Statistical Methods

Before we get into some of the statistical methodologies used for predictive analytics, there are two key things that need to be agreed upon by the internal management/team:

  • a common set of rules to govern the process – not enough to create additional bureaucracy, but enough to remove uncertainty and guesswork from the equation. This includes such things as how to collect and classify the data, how to manage the process, and how to ensure that certain biases are removed from the key steps in the process, as a lot of this is about gathering, analyzing, and synthesizing data
  • standardized timeframes – whether daily, monthly, or yearly, the timeframe needs to be agreed upon in order to keep consistency in the process

There are multiple statistical methods used to for Predictive Analytics, with three of the top ones being:

  • Moving Averages – a rudimentary way to analyze data, the Moving Average (MA) calculates next month’s demand as an average of the last 3 months. This can work for businesses who do not have a high degree of seasonality
  • Time Series Analysis – a more advanced analytical method that incorporates seasonality data, impact factors (related to the business ie. manufacturing delay), and quantitative data from experts in the business
Demand Forecasting, Resource Planning, and Procurement Strategy
  • Regression Analysis – is a statistical technique to analyze a dependent variable (target) and an independent variable (predictor) in order to find a causal relationship between the variables that affect a demand forecast. Depending on the complexity of the variables and the type of business, there are simple linear regression analysis all the way to more complex polynomial regression.

These are but a few of the possible statistical methods to forecast demand. There is no ‘silver bullet’ in terms of what type of analysis works best; different methodologies are optimized for different businesses. Having inventory or not, marketing strategies, reliance on seasonality, etc. are all factors that could drive a business from one methodology to another.

Predictive Analytics – Data Science

Part of the innovation, from a business modelling perspective, is that a lot of the more advanced techniques and statistical methods are becoming available now for anyone to use.

One such example is Facebook Prophet, which is available on Github as an open-source library for Python and R. Prophet is a predictive analytics tool that is based on the time-series method discussed above. Since Facebook developed it, it can handle large data sets and offer advanced analytics for those who are able to use Python or R.

And this is just the beginning. The ‘data science’ dimension of predictive analytics deals with the data itself. The initial pain point for most entities to build forecasting models is finding the data, cleaning it up, and then finding the appropriate model/methodology to build a forecast around.

But once the data is in some kind of model form, the sky is the limit, as there are multiple new and emerging data science applications and techniques – such as those that deal with advanced analytics or machine learning – which can be applied to any type of data set that pertains to business forecasting.

Overall, Predictive Analytics is a field that is greatly enhanced with easy-to-use software programs and open-source libraries for the purposes of modelling, forecasting, and planning. These methodologies and tools can help businesses forecast demand and manage supply, especially during an era of increasing instability and change.

Leave a Reply

Your email address will not be published.