Predictive analytics

Why it matters

Most expensive mistakes in an established business are timing mistakes: hiring too late for the busy season, ordering stock that sits for six months, discovering a customer has left after the renewal date. A forecast that is right within 10 or 15 per cent is enough to move those decisions weeks earlier. Prediction also scales attention. You cannot phone every customer, but you can phone the twenty the model says are most likely to leave this quarter.

Descriptive first, predictive second

The lead figure shows the four levels of analytics. Descriptive answers what happened; diagnostic answers why; predictive answers what is likely next; prescriptive answers what to do about it. The levels build on each other. A business that cannot reliably report last month's sales by product has no basis for forecasting next month's. A business that cannot explain why one site outperforms another cannot predict which new site will work.

This is why the honest advice to most SMBs is to get the first two levels right before investing in the third. It is also good news: by the time your reporting is clean and your definitions agreed, the simplest predictive methods become available in a spreadsheet.

A forecast you can build and check

The simplest useful forecast has three parts: a baseline (what a normal period looks like), a seasonal index (how much a given month or week usually differs from the baseline) and a trend (how fast the business is growing or shrinking year on year). Multiply them and you have an estimate that captures the two things most SMB demand depends on, season and growth. With two or three years of monthly data this takes an afternoon.

Its real value is that it can be checked. Each month, compare the forecast with what happened and record the error. If the forecast is consistently within 10 per cent, you can plan staff and stock from it. If it is wildly out in particular months, you have learned something about the business, such as a public holiday effect or a customer whose orders dominate.

When machine learning is worth it

Machine learning models earn their keep when the question has many inputs, when there are many things to predict, and when the prediction is used often. Scoring every customer for churn risk, estimating delivery times for every order, or predicting which equipment will fail from sensor and service history are good examples. Forecasting one business's monthly revenue usually is not; a seasonal model does nearly as well for a fraction of the effort.

Whatever the method, a model is only as good as the history it learns from and the care taken afterwards. Models drift as the business changes, need their accuracy measured against real outcomes, and need someone to own them. Building, deploying and monitoring models in a repeatable way is a discipline in its own right; SMBH Consulting's MLOps and AI application services exist for businesses that have reached that stage, but the first forecast should be yours, in a spreadsheet, so you understand what the model later has to beat.

Using predictions responsibly

A prediction is a probability, not a fact. Treat it as a reason to look closer, not a verdict. A customer flagged as likely to churn deserves a conversation, not an automatic price cut. A demand forecast informs a roster; it does not set it without a manager's judgement about local events. Keep humans in the loop for decisions that affect individual customers or staff, and be careful that models built on personal information comply with your privacy obligations.

Write down what each model is for, what data it uses, how accurate it has been and who reviews it. That record is what lets you trust the model next year, and what lets you switch it off the moment the business changes in a way the history cannot reflect.

How to start predicting

1. Pick a decision with a timing cost Rostering, stock orders, cash planning or renewals. The prediction must change what someone does.
2. Check your history At least two full years of clean, consistently defined data for the thing you want to predict.
3. Build the simple forecast Baseline, seasonal index and trend in a spreadsheet. Produce a range, not a point.
4. Use it for one real decision Staff the busy season or place the order from the forecast, with a manager's override.
5. Record the error every period Forecast against actual. After a few months you will know how far to trust it.
6. Add drivers if the error is large Weather, public holidays, marketing spend, key customers. Only move to machine learning when simple models stop improving.
7. Assign an owner and a review Someone checks accuracy quarterly and retires the model if the business changes.

Real-world example

A Brisbane air-conditioning business forecasts the summer rush

Northside Climate services and installs air-conditioning with 22 technicians. Every summer it is caught short in the first hot week of December and every winter it carries idle staff. The operations manager, Priya, pulls three years of weekly job counts from the job system. The off-season baseline is about 310 service jobs a week; the first weeks of December have historically run at around 1.9 times baseline; and the business has grown about 8 per cent a year. Her forecast for the first week of December is 310 × 1.9 × 1.08, roughly 636 jobs, with a range of about 580 to 690 based on past variation.

Each technician completes around 28 jobs a week, so 636 jobs needs about 23 technicians working flat out, with no allowance for leave or breakdowns. The business has 22. Priya hires two casual technicians for December, pre-books maintenance visits in October to pull demand forward, and tracks forecast against actual each week. The December forecast lands within 6 per cent. Encouraged, she builds a second, simpler prediction: units more than eight years old with two or more breakdowns in the last year are offered a replacement quote in spring, which smooths installation work into a quieter month.

Common mistakes

• Predicting before you can describe If last month's numbers are still argued about, a forecast of next month is built on sand. Get descriptive and diagnostic reporting right first.
• A point instead of a range "636 jobs" invites false confidence. Report the range from past error and plan for both ends.
• Never checking the forecast A prediction that is not compared with what happened cannot improve and should not be trusted. Record the error every period.
• Jumping to machine learning Complex models cost more to build and maintain and are often no better than a seasonal average for one business's demand. Earn the complexity.
• Letting the model decide Predictions are probabilities. Keep a human in the loop for anything that affects an individual customer or employee.

KPIs to track

• Forecast error The average percentage gap between forecast and actual. Within 10 to 15 per cent is useful for most SMB planning.
• Decisions made from forecasts Rosters, orders or campaigns that used the forecast. If none, the prediction is a report.
• Timing gain How many weeks earlier a decision is now made compared with before. The whole point of predicting.
• Model review cadence Whether each model was checked for accuracy and relevance on schedule. Quarterly is typical.

Frequently asked questions

How much data do I need to forecast?

For a seasonal forecast, at least two full years of consistently recorded data so that each season appears twice. For churn or failure models, enough examples of the thing you are predicting to find a pattern, which usually means hundreds of cases rather than dozens.

Is predictive analytics the same as AI?

They overlap. Simple forecasting is statistics in a spreadsheet. Machine learning, a branch of AI, is used when there are many inputs or many things to predict. Most small businesses should start with the simple end.

How accurate does a forecast need to be?

Accurate enough to change a decision for the better. A roster or stock order planned from a forecast within 10 to 15 per cent usually beats one planned from memory, and the error shrinks as you learn the drivers.

Can I predict which customers will leave?

Often, if you have a few years of history and a clear definition of "left". Simple signals such as falling order frequency or unresolved complaints go a long way before any model is needed.

Key takeaways

• Prediction builds on description: clean, agreed history comes first.
• Baseline × seasonal index × trend in a spreadsheet is a real forecast you can check.
• Report ranges, record the error every period and let accuracy earn trust.
• Machine learning is for many inputs and many predictions, not one business's monthly revenue.
• Predictions are probabilities; keep people in the loop and give every model an owner.