Forecasting Example

Our Forecasting as a Service provides a number of capabilities and benefits to a DIY solution:

• We constantly test & incorporate the most impactful forecasting methods
• We maintain the specialist infrastructure to optimise machine based forecasting - this is often significantly cheaper than general purpose AI/ML offered by Amazon / Google / Microsoft.  Where 3rd party services are required, we can integrate this with your service.
• Forecasting involves AI/ML forecasting technology combined with integration, visualisation, planning etc.  We maintain the whole stack for you - or integrate it with your existing tools.

• We understand the good forecasting involves machine + human intelligence and business acumen.  We provide tech, capabilities (coaching) and processes to enable your teams.

  Overview of Technical Forecasting Integration

1. We configure data feeds from your systems to our system to ingest your planning data.
2. We combine your data with datsets from other sources (economic data, price data, seasonality, confidence indexes, purchased category/consumer data e.g. Nielsen) and we assess to what degree each of these impacts your historical data (co-variates)
3. We detect features in your datasets that can be grouped together
   1. certain products may be complimentary or substitutes and have a positive or negative impact on each other
   2. certain product categories may respond differently to different co-variates. e.g. value lines might increase during the cost of living crisis
4. Anomalies will be detected and adjusted either from automated techniques or from human insight.  e.g. Trade promo windows may impact some products, you may experience supply issues that impact certain products or the timing of Easter may impact the timing of sales in March/April
5. Different time lags/warping will be analysed to understand timing impacts e.g. The price of Bananas may be a 3 month leading indicator of the price Cocoa which is a critical raw material in chocolate products.
6. We normalise / standardise your data to improve statistical performance with scaling 
7. We apply a range of AI/ML methods and evaluate which is most effective at different levels of your planning dimensions. e.g. Is it more accurate for forecast by customer and product by month or product by month.  In the testing we:
   1. Define a training & validation range of data which we use to automatically select the best model fit

   2. In evaluation, we also define a testing range which is not known to the model until after training and method is selected.  We then calculate the error of the selected model over the testing horizon. We do this in an automated way for all of your datapoints to select the best model for each combination.
       

      

      
      In the above chart, the black is the training dataset, the blue is the testing dataset and the system selected TiDE+RIN as the best fit method.  When we tested the purple testing (actuals) against the forecast, we see TiDE+RIN (orange) is visually a better fit than the next best two methods TiDE (cyan) and NHiTS (green).  The Mean Average Error (MAE) and Mean Squared Error (MSE) are in the table below:
       

      	mae	mse
      NHiTS	0.260374	0.083938
      TiDE	0.166268	0.038078
      TiDE+RIN	0.112519	0.021184

8. We evaluate the impact co-variates have had on the forecast and expose the impact on the forecast
9. We integrate this in to your systems so it can be further enhanced by humans

To discuss your specific requirements, please get in touch via our contact page.