M5 – are we forecasting for forecasting sake?

# #

Can we apply outcomes from the M5 forecasting competition to real world supply chains?

We, the Virtual Forecaster team at Hatmill, are enjoying pitting our wits against the world’s most eminent forecasters through the Global M5 competition. It is a competition open to all data scientists, mathematicians and forecasting gurus around the world to see who can create the most mathematically perfect forecast.

Are we forecasting for forecasting sake?

We are currently in the top 25% of the leader board using our Virtual Forecaster tools and techniques. However, taking part in M5 has made us wonder; “Are we forecasting for forecasting’s sake? Does this competition create better outcomes for Supply Chain to make better-informed decisions?” Can it really be applied in the real-world?

There are three key things we believe would make M5 more relevant to everyday scenarios:

1.      Products with very intermittent demand

2.      The forecasting time bucket used for the validation/evaluation

3.      The lack of formal communication around Promotions and Events

Products with intermittent demand and their Minimum Order Quantity

As mentioned in the M5 participant guide, a significant percentage of products have very intermittent demand (12.3% sell fewer than 1 unit per week and 40.5% fewer than 12 units every 4 weeks, based on an average of the most recent 52-week demand history provided).

Typically, Display Units (what is on the shelf) will contain more than 1 consumer unit (what the consumer buys; and therefore, what we forecast). At the same time, stores have strict Fronting and Facing policies and want maximum ‘on-shelf availability’. When you put those two together, forecasting items with such intermittent / low demand is meaningless as the replenishment process should be driven by a min / max approach.

This is highlighted further by the forecast submission allowing “fractional” units. To this day, I’ve never been able to visit a supermarket and buy 30% of a cabbage, or 7% of a bottle of wine.

Time bucket and store replenishment

Similarly, while some bulky items like toilet paper or beer will be delivered daily to stores, some less bulky items will be delivered in ‘weekly’ slots. This is done due to the display unit vs consumer unit argument (see previous point) and to avoid overwhelming store staff with restocking one item in one part of the store. Hence a weekly forecast would for some items be more than sufficient, rather than the unit-level daily forecast required in the M5 competition.

Promotions / events management communications

Finally, the whole of M5 is done without any interaction with Walmart commercial Team (M5 is using their dataset). This makes sense as it would be impossible to organise. However, a good Demand Planning process must have various inputs; such as:

  • New product introductions / delists;
  • Understand past performance deviations (and any history revision required);
  • Any promotional events and their nature (if the promoted item is different (50% free printed on-pack for example) and require a new SKU code or is it just a price promotion)

While competition is always healthy, but has it lost sight of why it is there? if it’s not fit-for-purpose and relevant to the real world, is it really helping?

What do you think?

By improving their demand forecast, we helped one client save £3m in agency staffing costs (that’s an 18% reduction).
What could we do for you?

Ideas & Insights

Sharing Our Expertise

Our guides, ideas and views. Explore our insights to deliver tangible improvements to your supply chain and logistics operations.