The methods ladder

Deep-dive lesson · about 12 minutes · short quiz at the end

ccPlanning academy · forecasting · deep dive

The methods ladder

Naive to Holt-Winters — and knowing which rung you actually need.

The principle

Climb only when the data asks.

Forecasting methods form a ladder, simple at the bottom, complex at the top. The skill isn’t reaching the top — it’s stopping at the lowest rung that captures the patterns actually in your data.

Each rung adds the ability to handle one more pattern.

The ladder at a glance

Four rungs, four jobs.

Each rung up handles one more thing the rung below ignores.

Rung 1

Naive

Tomorrow equals the last comparable period — this Monday equals last Monday. No maths, no model.

It sounds too simple, yet on stable, strongly seasonal data it is brutally hard to beat. It is also your benchmark: any fancier method has to earn its keep against it.

Rung 2

Moving average

Average the last n like-periods. It smooths out the random bounce so one odd day doesn’t whip your forecast around.

The cost: it lags. A genuine trend or step-change shows up late, because the average is still dragging old values along.

Rung 2 · the dial

Short vs long window.

A short window reacts fast but stays jumpy. A long window is smooth but sluggish. Choosing n is choosing how much you trust the latest data over the older data.

Rung 3

Exponential smoothing

Like a moving average, but recent periods get more weight and the weighting fades smoothly into the past — controlled by one parameter, alpha.

High alpha = react fast to the latest point. Low alpha = stay stable and ignore blips. It captures level beautifully, but plain SES still has no concept of trend or season.

Rung 4

Holt-Winters

Exponential smoothing extended with two more components: one for trend and one for seasonality. Three dials — level, trend, season.

Now the method can ride a growing baseline and repeat a yearly or weekly shape at the same time. For most contact-centre series, this is the top rung you’ll ever need.

Matching rung to data

Pick the rung that fits the patterns.

Flat and stable? Naive or a short moving average. A clear trend? You need a method that models trend. Strong weekly and yearly shape? Holt-Winters.

Using a rung that can’t see a pattern you have — or one that invents structure you don’t — both go wrong.

The overfitting trap

More complex is not more accurate.

A complex model with many parameters can fit your history almost perfectly — including its random noise — and then forecast the future worse than naive.

Fitting the past is not the goal. Predicting the future is.

How to choose, honestly

Test on data the model hasn’t seen.

Hold back the last few weeks, forecast them with each candidate, and compare against what actually happened. The rung that wins out-of-sample is the rung to use — not the one that looks cleverest.

Where machine learning fits

Higher rungs exist — they’re rarely the bottleneck.

ARIMA, gradient boosting, neural nets sit above Holt-Winters. They can help on large, complex, multi-driver problems — but only once your data is clean and the building blocks are right.

A sophisticated model on bad inputs is just confidently wrong, faster.

A worked choice

Which rung for this series?

Volume is up ~6% year-on-year, with a firm Monday-heavy weekly shape and a January peak. Naive copies last year but misses the growth; a moving average smooths the week but lags the trend.

Only a method that models trend and seasonality together — Holt-Winters — fits all three at once. That’s the rung the data is asking for: no higher, no lower.

The takeaway

Start at the bottom. Climb only when the data asks.

Benchmark against naive, add complexity one rung at a time, and prove each step out-of-sample. The best forecaster uses the simplest method that works — not the fanciest one available.

Now test yourself ↓

1 / 13

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

In depth: the lowest rung that works

Forecasting methods form a ladder — simple at the bottom, complex at the top — and the instinct to reach for the top rung is exactly backwards. The skill is to stop at the lowest rung that captures the patterns actually present in your data. Each rung up earns its place only by handling one more thing the rung below ignores; if your data doesn’t contain that thing, the extra complexity buys you nothing and can actively cost you accuracy.

What each rung adds

Naive — this Monday equals last Monday — needs no maths and is brutally hard to beat on stable, strongly seasonal data; it’s also your benchmark. Moving average smooths out the random bounce, at the cost of lagging genuine trends. Exponential smoothing does the same but weights recent periods more heavily, controlled by a single parameter, capturing the level beautifully — though plain smoothing still has no concept of trend or season. Holt-Winters adds exactly those two components, so it can ride a growing baseline and repeat a weekly or yearly shape at once. For most contact-centre series, that’s the highest rung you’ll ever need.

Why complex isn’t the same as accurate

A model with many parameters can fit your history almost perfectly — including its random noise — and then predict the future worse than naive. Fitting the past is not the goal; predicting the future is. The only honest way to choose is out-of-sample: hold back the last few weeks, forecast them with each candidate, and compare against what actually happened. The rung that wins on data the model hasn’t seen is the rung to use — not the one that looks cleverest in a slide. Higher rungs (ARIMA, gradient boosting, neural nets) do exist and can help on large, multi-driver problems, but only once the data is clean and the building blocks are right. A sophisticated model on bad inputs is just confidently wrong, faster.

The principle to remember: start at the bottom, benchmark against naive, add complexity one rung at a time, and prove every step out-of-sample. The best forecaster uses the simplest method that works.

The methods ladder