Hello Reader!

Should your forecast lie to your optimizer? It sounds absurd. But sit with it for a moment.

+ first briefing of the month --> find a personal update from me at the end!

The Forecast That Wasn't Wrong Enough

You've probably been there. The forecast looked reasonable. The model ran cleanly. Then demand came in stronger than expected, capacity hit its ceiling, and the business absorbed the hit: penalties, backlog, unhappy customers, maybe some lost accounts.

The natural, human conclusion is: "We need to forecast higher next time."

And here's the thing - that conclusion isn't stupid. It's intuitive, it's actionable, and it feels like it should work.

But there's another layer to this that makes it genuinely tricky. Even a perfect forecast doesn't guarantee good decisions when constraints are tight. Imagine demand exceeds your capacity and you forecasted it perfectly. You still face a hard question: which demand do you satisfy with what you have? Premium clients first, or whoever ordered first? A perfectly accurate number fed into a poor decision rule still produces a poor outcome. Input quality and decision quality are not the same thing.

Asymmetry Handling in "The Decision Factory"

In "The Decision Factory" (Adam deJans Jr. & John Brandon Elam, 2026), optimization consultant Evelyn is working with Fulcrum Logistics Inc. She lays out the core problem to her client Maya.

Fulcrum might decide to call in fifty drivers for the day based on that amount being able to handle expected demand. On a low demand day, they needed thirty-five. Fine - save fuel, pay a bit too many driver hours. On a high demand day, they need seventy. And they can't conjure twenty extra drivers. Above fifty, the overflow becomes backlog, expedited shipping, or lost sales.

Maya turns the forecast over in her mind. It's wrong about half the time in each direction - which is exactly what you want from an honest, unbiased forecast. So far, so reasonable.

Then Evelyn shows her the cost side. When Fulcrum calls in too few drivers, it pays thirty to forty times more than when it calls in too many. The cost of each is nowhere near symmetric.

The realization lands slowly. The forecast isn't the problem. The decision of how many drivers to call in - and how close to the limit to plan - is where the asymmetry actually bites. And the optimizer, doing exactly its job, pushes that plan right to the boundary of available capacity. The plan lives at the edge of the cliff.

The Fix Lives in the Optimizer, Not the Forecast

Three things worth taking from this.

First: keep your forecast honest. Train it on symmetric error metrics, evaluate it on symmetric error metrics. A forecast's job is to be as accurate as possible - not to compensate for how the downstream system behaves. The moment you start corrupting it with decision logic, you lose the ability to trust it for anything.

Second: the asymmetric cost belongs inside the optimizer. Tell your optimizer explicitly how much slack to leave between the recommended plan and the capacity ceiling. Make it a tunable parameter: the plan must stay at least X% below expected capacity. That is where the asymmetry logic lives - transparent, adjustable, and not buried invisibly inside a biased number.

Third: how do you find the right X? You simulate. Run the combined forecast-plus-optimizer system across hundreds or thousands of plausible demand scenarios. Find the slack level that minimizes total cost across that distribution. It is principled. It keeps every layer of your system doing what it is actually supposed to do.

So: should your forecast lie to your optimizer?

No. But your optimizer should know the truth about what a wrong forecast costs.

Over to You

Have you or your team ever nudged a forecast to compensate for what felt like an optimization problem? Hit reply - I'm curious how common this pattern really is.

Until the next iteration! Tim Varelmann ​Bluebird Optimization​ Complicated Decisions - Simply Automated! Follow me on LinkedIn​

​

PS: A personal update from Tim

Over Easter, I fell in love with a light social deduction card game called Saboteur. It is about a team of gnomes searching gold in a mine - with an unknown number of opponents, myriad ways to annoy your friends by placing utterly unnecessay obstacles in their way - plus at crunch time, a couple of bilevel optimization problems might have to be approximated in the player's heads. Check it out if your a game nerd like me! :)