Forecasting as a self-fulfilling prophecy
Every single week, a number of companies contact us asking whether Lokad could take care of their rolling weekly or monthly forecasts, say a few quarters ahead. Indeed, a decade ago, Lokad was founded around the idea of being a forecast as a service app. Unfortunately, this idea proved to be largely dysfunctional, and we moved on to better approaches. Some of the problems posed by forecasts are simple - deceptively so - but also fundamental. No matter how good the technology, it won’t do any good if its business premises are flawed.
Let’s be clear: Lokad does demand forecast, and predictive optimization remains our core competency, but there is a method to it; and the method starts by acknowledging that forecasts are largely a self-fulfilling prophecy, and that numerous feedback loops are involved.
Let’s illustrate this angle in more detail. If one seeks to forecast the sales of a fashion brand, one only needs to look at the quantities initially purchased: the quantities that get sold during the season are invariably equal to the quantities that were initially ordered from the suppliers (minus the shrinkage). Yet, frequently, the stock does not get sold at full-price, or conversely, the stock runs out halfway through the season. This supposedly “perfect” forecast is a mirage, and does not reflect the reality of the financial performance of the company.
The future demand is intrinsically coupled to future supply chain decisions. Yet, many companies are acting as if it were possible to forecast first (i.e. weekly demand) and then decide second (i.e. purchase order quantities). This approach is unwise because the latter profoundly impacts the former.
By way of anecdotal evidence, this coupling between future demand and future decision explains why so many “advanced” forecasting projects project fail: the backtest benchmarks were “proving” that the new forecast was better than the old one, and yet, once in production, things started to fall apart, seemingly for no reason. The reality is that, without due care, “dumb” forecasts tend to generate fewer adverse second order effects, i.e. unintended consequences caused by the forecasts themselves.
In practice, the retroactive feedback loop between forecasts and decisions take many forms:
- Introducing yet another reference into the assortment cannibalizes the rest of the assortment. Thus, making the demand forecast bigger for one reference should diminish (somehow) the forecast for all the competing products.
- The demand uplift resulting from a promotion for a given product heavily depends on the broader context. If the product happens to be the only one in the store to be promoted, the resulting uplift might be large. If every single product in the store happens to be promoted, the uplift is likely to be much thinner.
- Lead time should be forecasted, but the company might have a certain degree of control over the lead times if it can decide whether to ship the goods via sea or air freight. Both modalities require their own lead time forecasts; but the demand to be considered for the stock depends on the lead times.
- Buying larger quantities typically offers some kind of economies of scale, typically materialized through price breaks. A lower purchase price can then be turned into a lower selling price, boosting the demand so that it matches the larger quantity initially acquired or produced.
- Offering a discount to customers boosts the sales, but also modifies the future expectations of the clients. Customers will increasingly expect a discount, and will delay their purchasing decision until a discount is offered.
All those feedback loops are one of the key reasons why we, at Lokad, have become very uncomfortable to deliver “naked” forecasts. We have the firm intent to never repeat our early years’ mistakes, and in the spirit of the hippocratic oath, delivering value to our clients starts by not wreaking havoc in their supply chains through misguided approaches.
This is no accident if our Quantitative Supply Chain manifesto leans heavily on the decisions. It’s the consequences, sometimes far-reaching, of the decisions that need to be accurately predicted, not some kind of abstract demand. In this respect, our approach is aligned with the insights uncovered in the 1830’s by Jean Baptiste Say that supply creates its own demand.
Considering the present state-of-the-art progress in machine learning and software, there is no “packaged” way to tackle feedback loops through software. It takes a supply chain scientist with a keen understanding of the business challenges, as well as a healthy dose of discussion with seasoned practitioners, to figure out heuristics and models that are at least approximately correct in their capacity to put the predictive capacities of Lokad to a profitable use for the client.