Forecasting with AI: useful, with caveats

0:00 Forecasting with caveats

We'll sell forty-two hundred units next month. It sounds like a fact. It's actually a forecast, built on assumptions that nobody wrote down. AI assistants can now produce forecasts in seconds, from trend lines to established time-series methods, complete with charts that look authoritative. In this lecture you'll learn what forecasting can and can't do, the main approaches, why every forecast must beat a simple baseline, how to use prediction intervals and scenarios, the caveats to state every time, and a backtest you can run yourself.

0:37 Why it matters

Why does this matter? Because forecasts drive stock, staffing, budgets and targets. A forecast presented as a single number hides uncertainty, and decisions made on false certainty either over-order or under-deliver. And every forecast assumes the future resembles the past in specific ways. If those assumptions break, because of a new competitor, a price change or a moving holiday, the forecast breaks with them. Knowing that isn't pessimism. It's professionalism.

1:07 Main approaches

Here are the main approaches, conceptually. Naive and seasonal naive: next period equals the last period, or the same period last year. Moving averages and exponential smoothing, which weight recent data more and can handle trend and seasonality. Classical time-series models like the ARIMA family. Regression with drivers, like price, marketing spend or holidays. And machine learning or foundation models for time series, which can capture complex patterns when there's enough data, but aren't automatically better.

1:40 Beat the baseline

Here's the key idea: always compare any sophisticated forecast against a simple baseline on held-out data. Seasonal naive, same week last year, is surprisingly hard to beat for many series. Think of it like a new route suggested by a navigation app. Before you trust it, you'd want to know it's actually faster than the route you already know. If a clever model doesn't beat seasonal naive in a backtest, it isn't adding value, and the simpler method wins.

2:14 Prediction intervals

Now prediction intervals. A range like likely between thirty-seven hundred and forty-eight hundred supports better decisions than a single number: how much safety stock to hold, what budget contingency to keep. But beware: intervals from many methods are too narrow, because they assume the model is right and conditions are stable. So check coverage in your backtest. If far fewer than eighty percent of actuals fell inside your eighty percent intervals, widen them, or treat them as a minimum estimate of uncertainty.

2:50 Caveats every time

And caveats to state every time. Structural breaks: new competitors, price changes, regulation or platform algorithm changes can make history a poor guide. Limited history: a year or two barely captures seasonality. Feedback effects: a sales target changes sales effort. Data issues: stock-outs make sales data understate demand. And horizon: accuracy usually deteriorates the further ahead you forecast. Moving holidays deserve special mention in our markets. Ramadan and Eid shift by roughly eleven days each year, which breaks same-week-last-year logic.

3:25 A good forecast prompt

How should you ask the AI? Using Python on three years of weekly sales, forecast the next twelve weeks. Hold out the last twelve weeks, fit on the rest, and compare at least two methods against a seasonal-naive baseline using mean absolute error. Account for Ramadan, Eid and Black Friday, and tell me how. Give eighty and ninety-five percent prediction intervals. And list the assumptions and the events that would invalidate the forecast. Backtest, baseline, known events, intervals, assumptions: five ingredients.

4:00 Example 1: KSA gift product

First example, a business case. An online retailer in Saudi Arabia forecasts demand for a gift product ahead of the holiday season. The AI's first forecast uses a simple trend and misses that Ramadan shifted by about eleven days between years. After adding holiday calendar features and backtesting, the model beats seasonal naive modestly. The team orders to the upper part of the eighty percent interval for high-margin items and the middle for low-margin ones, a decision rule that uses the uncertainty instead of ignoring it. They also correct for last year's stock-out weeks.

4:41 Example 2: Karachi school supplies

Second example. A Karachi school-supplies wholesaler forecasts back-to-school demand. The first forecast misses that one year's school reopening moved by several weeks. The analyst adds a weeks to reopening feature, backtests against seasonal naive, and uses the upper part of the eighty percent interval for fast-moving items where stock-outs are costly. A scenario table, early, normal and late reopening, goes to purchasing, with a decision rule for each. Forecasting became a decision tool, not a number.

5:14 Watch me do it, part 1

Watch me run a backtest. I load weekly sales, set the frequency, and split: everything except the last twelve weeks for training, the last twelve for testing. The seasonal-naive forecast is simply the same week last year. Then I fit Holt-Winters exponential smoothing with additive trend and seasonality over fifty-two weeks, and forecast twelve weeks. I compute the mean absolute error for both. If Holt-Winters doesn't beat seasonal naive, I use the simpler method and say so.

5:47 Watch me do it, part 2

Then I check interval coverage. I ask for eighty percent intervals over the backtest period and count how many actual weeks fell inside. If only six of twelve did, the intervals are far too narrow, and I widen them before anyone uses them for stock decisions. Finally, I write the forecast up as scenarios with a decision rule, plus the list of assumptions and breakers, and I save the forecast so we can compare it with actuals later and learn from misses.

6:23 Using forecasts

And use forecasts in decisions properly. Pair them with scenarios, best, base and worst, and decision rules. Monitor forecast versus actual and update regularly. Record the forecast and its assumptions, so you learn from misses rather than rewriting history. And be especially cautious with AI-generated forecasts of market-wide or macroeconomic outcomes, which are far harder than forecasting your own operational series.

6:50 Measuring forecast quality

How do you measure forecasting quality over time? Track accuracy across many periods and items, not one headline number, because different products may need different methods. Keep a forecast log with the date, the method, the point forecast, the interval and the assumptions. Every month, compare with actuals, and ask whether misses came from bad assumptions, a structural break, or plain noise inside the interval. That habit improves your forecasts faster than any new model.

7:23 Common mistakes

Common mistakes. Presenting point forecasts without ranges. No baseline comparison. Ignoring moving holidays. Training on stock-out periods as if they showed true demand. Trusting intervals without checking coverage. And never comparing forecasts with what actually happened. Each of these is easy to prevent with the checks you've just seen, and each one, left unchecked, eventually reaches a decision-maker.

7:48 Recap and try this now

Recap. Forecasts assume the future resembles the past, so state the assumptions. Backtest on held-out data and always compare with a simple baseline like seasonal naive. Use prediction intervals and scenarios, and treat intervals as minimum uncertainty. Handle moving holidays, stock-outs and structural breaks, and monitor forecast versus actual. Try this now: forecast one of your time series with a twelve-period backtest, a seasonal-naive baseline, eighty percent intervals with a coverage check, and three written assumptions that would break it.