Predictive Analytics in New Zealand

Measure the forecast you already have

Every business forecasts something. Usually one person, in a spreadsheet, on the last Friday of the month, using judgement they could not fully explain.

That forecast is the thing to beat, and almost nobody knows how good it is — because nobody kept the old ones and compared them with what happened.

Before commissioning anything, go and find the last twelve monthly forecasts and compare each with the actual result. It is an afternoon's work and it produces the single most valuable number in the project: how wrong you currently are.

Two things follow. Sometimes the existing forecast turns out to be good, and the project is unnecessary. More often it is worse than anyone assumed, which makes the business case obvious.

Set the horizon from your lead times

Forecast only as far ahead as your decisions require.

• Supplier lead time six weeks? An eight-week horizon is what matters.
• Hiring takes three months? That is the planning horizon.
• Cash decisions made fortnightly? Forecast a quarter, review fortnightly.

Accuracy always falls with distance, so extending the horizon for its own sake makes the forecast look worse and helps nobody. Settle it from your lead times during quoting; it constrains the whole design.

Insist on a range, not a number

A forecast of "400 units" invites a decision as though 400 were a fact.

A forecast of "400, with an eighty percent chance of falling between 320 and 490" supports a genuinely different decision — and if stocking out costs you far more than carrying stock, you order towards the top of that range deliberately.

That is the whole point of forecasting well: it lets you size the risk rather than pretend it away. Ask each firm whether their forecasts carry prediction intervals. A supplier delivering point estimates only is making the output look more certain than it is.

Where forecasting works, and where it does not

Be realistic about your own data before spending anything.

Forecasts well: steady demand with seasonality, subscription revenue, staffing against predictable patterns, payment behaviour on a large customer base, anything driven by weather where the weather itself is forecastable.

Forecasts poorly: demand dominated by a few large customers whose decisions are made in a meeting you are not in; new products with no history; anything driven by a tender or a contract award; markets that have just changed structurally.

If you are in the second group, no model fixes it, and a firm that says so is being honest rather than unhelpful. The better project there is often to capture the leading indicators — quote pipeline, enquiry volume, customer intentions — so that in a year there is something to forecast from.

Decide what changes because of it

Name the decision, the person and the change, in the brief:

• "The purchasing manager sets reorder quantities weekly instead of monthly."
• "Credit control chases the top twenty likely-late invoices on the first."
• "The account manager calls the ten highest-risk customers each week."

A forecast that does not change what anyone does is a chart. It will be admired for a month and ignored by the second quarter. This is the most common way these projects fail, and it is entirely predictable in advance.

Track accuracy from day one

Insist this is built in rather than added later:

• Every forecast is stored as it was made, not overwritten.
• Actuals are matched against it automatically.
• Error is reported in a measure the business understands — units, dollars, percentage — not only in statistical terms.
• The naive baseline is tracked alongside, so you can always see whether the model is still earning its place.

Without this, within six months nobody can say whether the forecast is any good, so it becomes a matter of opinion and then a matter of neglect.

Simple methods are often enough

For a lot of New Zealand businesses, a well-implemented seasonal model on clean data delivers most of the available benefit. It is quick to build, inexpensive, easy to explain to a board, and easy to correct when it is wrong.

More sophisticated approaches earn their cost when you have many series to forecast at once, when external drivers genuinely matter, or when the simple method has been measured and found wanting.

Ask each firm what they would try first. Anyone whose first proposal is the most complicated option available should be asked what the simple version scored.

Comparing five quotes

Read for:

• Who asked how accurate your current forecast is?
• Who set the horizon from your lead times?
• Who offered prediction intervals rather than single numbers?
• Who asked which decision changes, and who makes it?
• Who included accuracy tracking from the start?

Describe what you are trying to forecast, how far ahead you need it, what decision it feeds, and how much history you hold. Firms who do this work will ask to see your last year of forecasts against actuals before they promise anything.