SCM Awareness: The Forecasting Process 2
Continuing on our look into the Forecasting process now we will look at some forecasting methods.
A key decision you will have to make is Choosing what method you will use. Qualitative or quantitative or most likely a mix of both.
Quantitative methods look at historical data. Qualitative methods look at opinion. This can be the most daunting stage of the process. Many people think they need a Doctorate in Mathematics to be able to use statistics. However, the forecasting tools that exist will do the calculations for you. Even if you are just using excel for your forecast there are tools within excel that can help you. And finally, no matter what method is used the key final calculation in the forecast will be done in the head of the forecast owner simply by asking two questions
- What do I think the forecast for this item will be?
- Am I happy with those numbers?
Human brains have immense data processing power. Even the best computer will struggle to match a trained, well informed, and confident Forecast owner. This is known as qualitative methodology. So even if you cannot invest in a very expensive top of the range forecasting system you can still achieve a very accurate forecast with the resources at your disposal.
Quantitative methods. Quantitative methods use formulas or algorithms to calculate the forecast. A Forecast Algorithm is the calculation or set of rules used to develop the forecast. Forecasting is a mathematical discipline in itself and we won’t go into it in detail here, but I will give a brief overview of some methods you may come across.
There are two main methods. Time Series Forecasting and Associative Forecasting.
Time Series Forecasting is more common because it is easier to calculate and is also easy to understand in its approach. It assumes that trends and patterns in the past will continue into the future. Elements such as seasonality, business performance, economic cycles etc can all be factored into the calculation. Time Series forecasting is relatively straightforward and a lot of the mathematics you will have studied in school or college will cover this.
Simple Moving average. This is where you take the average of the sales over a set period. Say you are in June and want to forecast for July. You do not yet have the sales numbers for June. The policy is to take the average of the last 4 months of sales. In that case you would take February March April and May and use that for the forecast in July. In a case like this you may even take that number and apply it for all months in the forecast horizon. In a month’s time you now have the sales data for June, so you move everything on a month. The forecast for August etc now becomes the average of March, April, May and June.
Weighted Moving Average. This is similar in concept to the simple moving average, but you apply more importance or weight to the more recent results. The weights are often decimals or percentages but must add up to 1 or 100%. Continuing on the previous example You might decide to give the newest month a weighting of 50%, 25% for the next back, 15% for the next and 10% for the next. You multiply the sales in each month by their weighting and then add the four results together to get the total. The idea here is the forecast gives you more of a chance to react to demand trends.
Exponential smoothing: Is similar to Weighted Moving average in that more recent results get a higher weighting but is different in how it is calculated. Here you use both the actual demand and the forecasted demand from past time periods and weight them against a smoothing factor. The calculation reduces the weighting of each past month by the smoothing factor. Let’s say 25%. The first month gets a priority of 75%. The next gets a priority of 75% of 75%. The next gets a priority of 75% of 75% of 75%. This sounds like a lot of maths but it is actually quite straightforward to do in excel with a couple of formulas. The trick is to set the forecast of first month in the historical data to equal the demand of that month and then the formula will work for all future months. This is one of those concepts that is straightforward when you get it but until you do then it can be quite confusing. Don’t worry too much about it as it is easy to do in excel and if you have a forecasting system it will probably do this for you.
Associative Forecasting is more complex in its mathematics and in the input data that is required to be converted into numbers first to help the calculation. The two most common forms of Associative Forecasting are
Linear Regression. Here you calculate the forecast based on a historical predictor value (e.g. last month’s sales combined with the slope of the trend and the point where that slope reaches 0. The difficulty most people have is in calculating these elements as the formula for the forecast itself is then relatively straightforward (Forecast = Point there the slope hits 0 + (The slope multiplied by the historical predictor value). But calculating the slope, predictor value etc is difficult.
Multiple Regression. This is like linear regression except it adds in harder to calculate variables. You need to look at factors that seem to influence the forecast and decide if they are “leading indicators” or “lagging indicators”. Leading indicators are events that you know have happened and generally then lead to a change what will happen in your forecast (e.g. if you make covers for iPhones then a change in sales of iPhones should see a similar change in your sales in the future) Lagging indicators are events that you know have happened but indicate that a change has already happened in your forecast (for example if your inventory has gone down below your desired level and there was no problem with production etc then that may mean sales were higher than expected in the past).
It is important when looking at these elements that you keep in mind the difference between causation and correlation. Causation is something that has been directly caused by the event you are looking at. Correlation is something that happened at the same time but was not caused by it. For example, the weather was good and you saw sales increase. That makes sense if you produce something like ice cream. However, if you produce baked beans then there is no obvious reason why an increase in sales would result from good weather. There might be a reason and you should investigate it to see if there is, but it is most likely a coincidence and thus is correlation. Another well knowing example is that Divorce rates in the US state of Maine between the years of 2000 and 2010 was found to track very closely to the consumption of margarine in the state over the same time. This is known as Spurious Correlation.
So, you can quickly see that effort has to go into figuring out the supporting data in associative forecasting. It can give you more accurate results, but you have to work out if it is worth the extra effort. You do not often see this in DIY excel files and it is in more complex dedicated forecasting systems. But if you are really struggling with these statistics and want to look up some more there are tons of resources online.
Qualitative methods. Qualitative methods look at personal opinion and expertise. This relies on the judgement of the forecast owner or key members of the business to develop the forecast. You have to be aware of bias (e.g. if a salesperson is always very optimistic) however even with that this is probably the method you will use more than quantitative methods. This is because business owners and experts should know more about the business than any system. They will intrinsically know the trends and cycles, the threats and opportunities, and the competitive environment. The forecast owner should be integrated into this executive ecosystem and trusted with the data they need. The sort of information you would look at here include:
- Historical review
- Conversations with Sales Teams.
- Personal opinion
- Market analysis
- Competitor analysis
- Customer Surveys
There are ways to help reduce the impact of bias in the forecast. One technique is to ask the sales team to provide a worst case, most likely case and best case set of numbers. This allows the salesperson to get their bias covered while at the same time also giving you what their experience tells them is the most likely outcome. But you may want to hedge your bets a little and consider the low and high estimates. One common recommended approach to this is to add the low and high estimates with 4 times the most likely and divide by 6 to get the new number. Another approach is to have a team of people review the data and come up with an agreed number for the forecast. This is a sort of wisdom of crowds approach. Of course, the “crowd” will most likely only be 3 or 4 people but having a team discuss and trash out the details has been proven to come up with a more accurate number. And it does not have to take very long. In my experience if the data you have gathered in advance is accurate and comprehensive the team will agree very quickly on most items.
A combination is best I have yet to see a forecasting methodology that was perfect in itself. I have seen some very good calculations, a former member of my team in one company did his masters thesis on forecasting and came up with some very good formulas that we used in an excel tool. They were complex and excellent formulas but even with all his work they still were reviewed by the team and adjusted if necessary. In my experience there is no quantitative formula driven forecast that was not made better after being reviewed and discussed by the team. Likewise, there is no qualitative human centered approach that was not greatly assisted by some mathematics to back up the assumptions. If you go into the meeting with an estimated forecast on a spreadsheet that is the foundation of your discussion. That is why I would advise you not to put all your eggs in one basket. Use a few different methods working together to come up with your forecast.