Machine learning in forecasting

Deep-dive lesson · about 10 minutes · short quiz at the end

ccPlanning academy · advanced

Machine learning in forecasting

The top of the methods ladder — powerful, hungry, and easy to misuse.

The big idea

ML is another rung, not a magic wand.

Machine learning — gradient boosting, neural nets, the like — sits above Holt-Winters on the methods ladder. It can capture patterns classical methods miss. But it has to beat naive and Holt-Winters out-of-sample to earn its place, exactly like any other method. Sophistication isn’t accuracy.

What ML adds

It learns complex, multi-driver patterns.

Classical methods extrapolate a series’ own history. ML can blend many drivers at once — weather, marketing spend, web traffic, prior-day spillover, holidays, price changes — and find non-linear interactions between them. Where demand genuinely depends on many external signals, that’s a real edge.

When it genuinely helps

Lots of data, lots of drivers, real complexity.

ML earns its keep on high-volume operations with long, clean histories and rich driver data — and on problems where the relationships are too complex to hand-craft. If you’re a 30-seat centre with two years of monthly totals, Holt-Winters will likely match it for a fraction of the effort.

Trap 1

Data hunger.

ML models need lots of clean, well-labelled history to learn from — and rich driver data to be worth it. Most contact centres have less usable history than they think. Feed a hungry model thin or dirty data and it learns noise, not signal.

Trap 2

Overfitting.

A flexible model can fit the past almost perfectly — including its random noise — then forecast the future worse than naive. The methods-ladder warning applies double here: always judge ML on data it has never seen, never on how well it explains history.

Trap 3

Opacity.

A Holt-Winters forecast you can explain to a sceptical ops director. A neural net’s output is hard to justify — “the model says so” doesn’t survive challenge. When you can’t explain why the forecast moved, you can’t defend it, adjust it sensibly, or trust it when it surprises you.

How to adopt it sensibly

Benchmark, hold out, monitor.

Run ML alongside your current method, compare both out-of-sample over real periods, and only switch where it consistently wins. Keep measuring forecast value add after you adopt — an ML model that drifts can quietly become worse than the naive baseline it replaced.

The honest position

Most centres don’t need it — yet.

Clean data and well-built classical forecasts beat fancy models on bad inputs every time. Fix the fundamentals first — the building blocks, clean history, forecasting AHT and shrinkage. ML is the last 5%, not the first move. When the fundamentals are solid and the scale justifies it, it can be a genuine edge.

The honest benchmark

Does the clever model actually win?

Run the neural net and Holt-Winters and naive on the same held-out weeks. ML fits last year beautifully — 2% error on history. But out-of-sample it lands at 9%, Holt-Winters at 8%, naive at 11%. The fancy model lost to the simple one it was meant to beat.

That’s overfitting: it memorised noise. The only verdict that counts is on data it has never seen — and here, sophistication didn’t earn its place.

The takeaway

Earn-its-keep, like every other rung.

ML can blend many drivers and find complex patterns — but it’s data-hungry, overfits easily and is hard to explain. Benchmark it out-of-sample, adopt only where it consistently wins, and never let it skip the fundamentals. Sophistication has to prove itself, same as naive does.

Now test yourself ↓

1 / 10

Slides done? Here’s the same idea in a bit more depth — the part worth keeping.

In depth: powerful, hungry, and easy to misuse

Machine learning — gradient boosting, neural nets and the like — sits above Holt-Winters on the methods ladder, and it can capture patterns classical methods miss. But it’s another rung, not a magic wand: it has to beat naive and Holt-Winters out-of-sample to earn its place, exactly like any other method. Sophistication isn’t accuracy, and the model that explains history best is often not the one that predicts the future best.

What it adds, and when it helps

Classical methods extrapolate a series’ own history; ML can blend many drivers at once — weather, marketing spend, web traffic, prior-day spillover, holidays, price changes — and find non-linear interactions between them. Where demand genuinely depends on many external signals, that’s a real edge. But it earns its keep specifically on high-volume operations with long, clean histories and rich driver data, and on problems too complex to hand-craft. A 30-seat centre with two years of monthly totals will likely see Holt-Winters match it for a fraction of the effort.

Three traps, and the honest position

The traps are real: data hunger (most centres have less usable history than they think, and a hungry model fed thin or dirty data learns noise), overfitting (a flexible model fits the past’s noise and then forecasts worse than naive, so judge it only on data it has never seen), and opacity (“the model says so” doesn’t survive a sceptical ops director, and a forecast you can’t explain you can’t defend, adjust or trust when it surprises you). Adopt it sensibly by running it alongside your current method, comparing out-of-sample, switching only where it consistently wins, and continuing to measure forecast value add afterwards. The honest position is that most centres don’t need it yet: clean data and well-built classical forecasts beat fancy models on bad inputs every time, so fix the fundamentals first — ML is the last 5%, not the first move.

The principle to remember: ML has to earn its keep like every other rung. It can blend drivers and find complex patterns, but it’s data-hungry, overfits easily and is hard to explain — benchmark it out-of-sample, adopt only where it consistently wins, and never let it skip the fundamentals.

Machine learning in forecasting