Managing inventory during and after crisis part 1 : the machine learning difference.

Ifmaking inventory decisions felt like casino-gambling during confinement, the after may be even harder… or not. Let’s share our experience – part 1 :

1. Using dedicated additional external explanatory data for forecasting/buffer dimensioning:

When you have some physical points of sale or warehouses and/or operate on short lead times, our data scientists indeed could find helpful explanatory data: indexes like the Oxford Stringency Index or Google Mobility have proven very valuable indicators and well-correlated with demand variations. Though certainly helpful, as far as our experience is concerned we found this solution far from being a panacea.

Pluses  :

• Can bring good forecasting power for short lead times and decisions (a few days in advance) and at aggregate level
• Can give some understandable explanations and feeling of control to decision makers

Minuses :

• Can introduce noise and unneeded complexity to the model at part number level
• Requires to forecast the explanatory data value itself for mid and long term decisions.
• Not really helpful in the early days of the crisis: reliably introducing fresh explanatory data in the model takes some days even if you are very focused (maybe you were smart or lucky enough to use this variable before crisis… we were not).

Some more explanations

Model complexity and displaced correlation: Nadii found the above explanatory data very powerful at the shop/warehouse/factory level, and algorithms using it gained a lot of weight in Nadii decisions. At part number level, it really depends on the product. Even famous the toilet paper peak demand, for example, is correlated with the rise of the index but displaced in time (some days before… so in this case the index explains the past rather than future). As usual external data can help, but users should not expect magic…

The Nadii optimisation engine, using forecast accuracy measurements, proved that the impact of these algorithms at part number level is really valuable for some items, but overall not better than simpler algorithms using the shop/warehouse trend (itself using explanatory data).

Limited power for longer horizons and forecasting lead times: as usual with explanatory data, it is easy to explain the past, but what we expect is prediction. Stringency index as a delayed effect on some days in the future, but if you need to anticipate more than some days ahead, you need to forecast the index itself, which is only displacing the problem. The effect is even worse on lead times, because you have to forecast lead times even more in advance.

2. Applying strategies used for new/early-stage products:

For products that started behaving completely abnormally and cannot be fully explained via Covid-specific explanatory data, Nadii saw a loss in accuracy for history-based or usual feature-based algorithms. Nadii then could find help in the algorithms used for new products:

1. Algorithms looking for similarities and forecasts within the product family or location.
2. Algorithms focused on short term trend identification and reactivity,
3. Algorithms using user-provided information (budget forecasts, confirmed end-customer orders…)

Pluses :

• Robust and simple alternative that Nadii can turn on when detecting a loss in late accuracy

Minuses :

• This is an OK strategy as long as short term events have much more impact than usual features or historical information (so not enough for highly seasonal demand or weather/vacation-dependent products). The system must keep an eye on more usual strategies to take over when situation returns to normality.
• If your system does not use machine-learning driven auto-adaptation, such behaviour must be turned on/off manually during and after an exceptional event, not realistic at part number level.
• If your system does use machine-learning driven auto-adaptation, data cleaning should be used to help accuracy when effects of the black-swan event disappear.

3. Cleaning Lead times and demand history :

We’ve noticed that for a large part of the product portfolio of our customers, some of the usual algorithms (statistical or machine learning without Covid features) still did well at spotting changing trends and show good performance during exceptional events or after the first shock passed. No need to panic, the core of Nadii did a very good job after this shock by isolating algorithms that did not perform well and promote the most adapted to new circumstances. For these usual algorithms and products though, late trends might still be noisy.

Although the years 2020 and 2021 will most probably be different from past years, a lot of information like seasonality or trend changes can be found in history. Auto-adaptation (see next blog article) performs the sorting between obsolete and relevant information for us. That’s why history is not irrelevant at all for our customers: we keep this precious data, but need to clean it to maintain high accuracy, only for algorithms and products where additional explanatory data was not enough to learn explain during confinement.

Pluses :

• Enables to keep part of the still-relevant historical information (like seasonality) and avoid post-confinement forecasting or errors.

Minuses :

• Needs to be automated (don’t count on overwhelmed users to do it for you).
• The rise in variability and uncertainty is very valuable information (see separate part 2 of this article) that can be lost when cleaning. Particularly true for the part of your inventory dedicated to protect from variations.

More detailed explanation :

a. Why perform cleaning ?

Let’s take an example: if you have a seasonal peak around March (e.g. better weather), and this year your sales were up 50% due to coronavirus at this same period, algorithms which do not grasp the Coronavirus effect may conclude that we have a higher trend this year, and maybe propose to continue with this higher trend in April or even next year at the same period. If we assume it is a one-off incident, we should therefore replace this peak by some reliable forecast right before the crisis.

b. What to clean?

We saw positive effect of data cleaning in 2 main areas :

1. Demand and lead times peaks : For very impacted products (e.g. our famous toilet paper), chaotic supplier lead times and demand history that could not be linked to explanatory data should be cleaned to avoid impacting your forecasts or buffer dimensions when back to more usual conditions.
2. Stockouts: if you unfortunately had stockouts, your system needs to correct demand automatically (taking into account lower sales AND the potential artificial peak when availability is back) if you do not want these algorithms to predict similar behaviour in the future.