When looking at a single product at a time, there is simply not enough data to produce an accurate statistical forecast. Indeed, on most consumer markets, the lifecycle of a product is less than 4 years, which means that, on average, most products don't even have 2 years of history available - that is, the minimal depth to perform a reliable seasonality analysis when looking at a single time-series.
We address the problem through statistical correlations: the information obtained on one product helps to refine the forecast of another product. For example, Lokad autodetects the applicable seasonality for a product even if the product has only been sold for 3 months. Indeed, while no seasonality can be observed with only 3 months of data, if older longer-lived products are present in the history, then the seasonality can be extracted there and applied on newer products.
While leveraging correlations within the historical data vastly improves the accuracy, it does increase the amount of computations to be performed as well. For example, to correlate 1,000 products looking at all possible pairs, there are a bit less than 1,000,000 combinations. Worse, many companies have a lot more than 1,000 products.
By leveraging cloud computing, when clients push their data to us, we allocate the machines just when we need them; then, less than 60min later on, we return the results while we deallocate the machines accordingly. Since the cloud we use (Microsoft Azure) is charging us by the minute, we only consume the capacity that we really need. As no company needs to forecast more than once per day, this strategy cuts hardware cost by more than than 24x compared to traditional approaches.
The traditional forecast is a median forecast, that is, a value that has 50% chance to be above or below future demand. Unfortunately, this classic vision does not address the core concerns of commerce: avoiding stock-outs and reducing inventory.
In 2016, Lokad introduced the notion of probabilistic forecasts for supply chain where the respective probabilities of every level of future demand get estimated. Instead of predicting on value per product, Lokad predicts the entire probability distribution.
Probabilistic forecasts vastly outperform classic forecasts for slow movers, erratic sales and spiky demand. We believe that 10 years from now, all companies serious about inventory optimization will have gone probabilistic probably leveraging a descendant of this technology.
Our clients send us data, typically through flat files, sometimes through databases, and we return results. The forecasts are provided as a service. No statistical skill are expected from our clients as Lokad is managing the entire process.
There is no statistical setup involved. Once the data is pushed in the proper format - no data cleaning required - Lokad returns the results in less than 60mins. It does not matter if it's the 1st time or the 10th time that the data is pushed to Lokad, our forecasting engine is fully robotized and require no manual intervention.
We have a large library of statistical models. It includes well-known classics such as Box-Jenkins, exponential smoothing, autoregressive and all their variants. However, classic models poorly leverage correlations. Thus, we have developed better models that take advantage of all the data made available to us.
Since the very beginning, we have been continuously monitoring the forecasts we deliver. Every day, we run forecast simulations to carefully assess the remaining weaknesses of our technology. Those findings help us to focus our development efforts where it matters most. Our clients benefit from an ever improving technology.