You are curious about how to use statistics and machine learning to build a demand forecast? Before jumping in, why don't you read first our master post on how to make a sales forecast?
The old way to build a sales forecast
Excel, the one and only tool.
When you want to build a sales forecast, the fastest approach would be to open an Excel spreadsheet, download all the historical values you have, look at them for quite a while and start building a model, maybing by building a pivot table and using a couple of formulas.
You would generally then start by indentifying the mains drivers of the business, if not already identified, such as number of active customer making a transaction per month, average value of the transaction, amount purchased, add many other drivers which belong to your business and then replicate it to all the regions where you operate in.
Year over Year trends and business knowledge
Next step would be to project the past trend to the coming months, generally using the best information you have, such as year over year growth observed and ... your knowledge of the business. E.g. which seasonality is being observed? Would you expect a sudden burst in market share following the launch of a new product? Contribution to growth analysis and CAGR can be helpful to get your head around the growth and internal patterns.
The period you would look at would be typically the monthly one. Looking at seasonality fits the financial calendar where months and month to date are generally the north star time dimension.
You might go to a weekly view - but almost never build a daily forecast; that would be way to daunting and one might get lost into details very quickly, losing the big picture.
Building a forecast with statistics
If you have a single small ad hoc forecast, doing it the traditional way is fine. When you need however to build many of them, let's say 20, this might be very exhausting and repetitive to be tackled manually. If you have multiple dimensions to forecast, let's say a forecast for each collection/season, product family ("trouser children") or per customer (assuming you want to project the activity of your top 50 customers); then you need to have a statisctically driven forecast, which will speed your manual reviewing. You won't have to start from scratch if you use a statistical forecast.
Also, for those who have been in the budget planification process, we usually have 2 processes being run in parallel:
- A Top down process - driven by the senior management
- A Bottom up process - kicked off and driven by the Sales department
Running multiple bottom-ups needs is a huge task and requires some kind of proxy, heuristic, allowing to go fast: this is where statistical forecasting models kick in. At the era of machine learning and data automation, many time series algorithm model exist.... but still require the expertise of statisticians or data analysts with advanced IT / programming skills.
Hire experts... or use datainsightout!
Large companies have this kind of profile and can attract them - but middle ones are struggling to get them. Hiring a data scientist to do a forecast is somewhat overrated, a data analyst could do it - provided he has the right skills in terms of manipulating programming languages such as python or R; and an IT developper... might not related with it.
This is where DataInsightOut might help you.
The solution allows you to run easily, just by uploading a dataset, multiple forecasts, using the Prophet algorithm, developped by the core data science team at Facebook. It a safe and open-source algorithm, developped, precisely to allow users to forecast out-of-the-box. DataInsightOut is making its usage easy, for non-tech people to use it.