Is maximum #supplychain Resilience one of your top 3 priorities for 2023?
Here’s a thread exploring what your primary tool should be
Decision models vs Forecasting models vs Visibility solutions.
Let’s start with pros & cons of each, the tradeoffs, and how they may interact–
Here’s a thread exploring what your primary tool should be
Decision models vs Forecasting models vs Visibility solutions.
Let’s start with pros & cons of each, the tradeoffs, and how they may interact–
1/11 Starting with visibility.
No one disputes that visibility is a necessary 1st step. You do need to see where your stuff is, where the blockages & delays are. But the issue is: does that really offer any solution to mitigate those disruptions? The straight honest answer is NO!
No one disputes that visibility is a necessary 1st step. You do need to see where your stuff is, where the blockages & delays are. But the issue is: does that really offer any solution to mitigate those disruptions? The straight honest answer is NO!
2/11 Visibility is a monitoring tool, it’s a great stethoscope, or X-ray device! Recall the Evergiven which got stuck in Suez Canal for 6d in ‘21 Mar blocking 369 ships behind it! It was very visible to all from day one, even from the Intl Space Station!
3/11 But what were the risk-mitigation actions offered by the visibility tool? None!
It’s not in its scope to recommend such actions to those companies who had containers on those stuck ships with raw materials heading to factories or finished products heading to DCs/retailers!
It’s not in its scope to recommend such actions to those companies who had containers on those stuck ships with raw materials heading to factories or finished products heading to DCs/retailers!
4/11 So visibility has a role to play but it can’t be the primary tool to deliver resilience.
Next we consider forecasting models, surely they can be helpful. Yes, absolutely if every year looked like 2019 where every link in the value chain was in a normal scenario …
Next we consider forecasting models, surely they can be helpful. Yes, absolutely if every year looked like 2019 where every link in the value chain was in a normal scenario …
5/11 ..there were no COVID lockdowns, no factory shutdowns, no spikey lumpy demand, no bull whip effects in consumer spend etc.
Then yes, a model trained on historical data from Jan 2017–Dec 2019 would perform w high accuracy in 2020 if none of the above shocks took place!
Then yes, a model trained on historical data from Jan 2017–Dec 2019 would perform w high accuracy in 2020 if none of the above shocks took place!
6/11. what happened by Mar 2020 had 0 correlation with Q4 of 2019 and any forecasting model trained on that historical data was seeing completely different actuals to act on. That was the issue every planning team faced through 2020 up to Q1 of 2022. Accuracy faded by 25%+
7/11 It became absolutely clear to every supply chain planner during those 2y--a forecasting model trained on a normal scenario data-set would be useless within a shocked environment! So how do we solve this if all we have is forecasting?
Well you have 2 approaches both painful
Well you have 2 approaches both painful
8/11 approach (a) Mix data from normal scenario 2019 and pandemic period data ( 2020-03--present). However, that seriously confused such forecasting models b/c the patterns were dramatically different.
b) You can train tens or 100s of models for each data-set for each mkt!
b) You can train tens or 100s of models for each data-set for each mkt!
9/11 If we have a perpetually shocked world as in COVID, then port delays, then Mississippi river barge volume decline, then Ukraine war, etc. Then you’d need one forecasting models for normal, then one for shock-A, shock-B, Shock-C etc. that's why forecasting has been failing!
10/11 Enter AI Decision Models-
Instead of tens of fcstg models, you simply have one monolithic decision model w multiple AI agents that are trained on each scenario to optimize for your weighted KPIs. The planner simply chooses the recommended action that fits current scenario!
Instead of tens of fcstg models, you simply have one monolithic decision model w multiple AI agents that are trained on each scenario to optimize for your weighted KPIs. The planner simply chooses the recommended action that fits current scenario!
11/11. We may get a new shock in 23 H2 or contemplate a shock for '24, we can simulate it now as a new scenario in the model's env (digital twin) as shock_E and we'd have an AI decision agent optimizing for KPIs when it occurs. That's the only way to achieve true resilience!
• • •
Missing some Tweet in this thread? You can try to
force a refresh
