Time-series forecasting: seeing demand before it happens

Most Icelandic companies are sitting on years of sales history in their systems but use it almost exclusively to look in the rear-view mirror. Time-series forecasting turns this around: it uses the history to predict the coming weeks and months. This requires neither enormous amounts of data nor magic, but a structured method, a simple baseline and an honest measurement of whether the forecast is actually better than the most obvious guess.

What is a time series and how does it become a forecast?

A time series is simply a sequence of measurements in chronological order: daily sales, weekly web traffic, monthly revenue. What makes time series special is that the order matters. Today's sales are related to yesterday's, and sales in December are related to sales last December.

A forecasting model takes advantage of this by breaking the series down into three parts. In its simplest form it can be written like this:

$$y_t = \text{trend} + \text{seasonality} + \text{noise}$$

The trend is the long-term direction: is the business growing, standing still or shrinking? Seasonality is the recurring pattern that comes back again and again at a fixed rhythm: weekly, monthly, yearly. The noise is everything else, random fluctuations that no model can predict. A good forecast captures the trend and the seasonality and acknowledges the noise as uncertainty instead of pretending to see a pattern in it.

Icelandic business is seasonal at its core

Few economies swing as regularly as Iceland's. The Christmas season is decisive in retail, summer traffic in restaurants and tourism, and the tourist cycles reach into everything from car rentals to bakeries. A company that plans purchasing, staffing or marketing as if every month were the same is paying for it, either in stock-outs and understaffing or in shrinkage and overstaffing.

On top of the annual rhythm come external variables that move demand from day to day. The weather drives sales of everything from ice cream and barbecue meat to outdoor clothing. Paydays around the turn of the month show up clearly in card turnover. Public holidays, trade fairs and the start of the school year leave their own marks. The advantage of these variables is that many of them are known in advance: we know when the month ends, and the weather forecast for the coming days is available. A model that receives this information has an edge over a model that only sees the sales themselves.

The honest workflow: baseline, backtesting, retraining

The most common mistakes in forecasting are not technical but methodological: jumping straight into a complex model and believing a pretty line on the screen. An honest workflow looks like this:

Start with a simple baseline. The simplest forecast is to say that next week will be like the same week last year, or like last week. This sounds naive but is surprisingly hard to beat in a seasonal business. A model only deserves credit for the part it does better than the baseline, not for the forecast as a whole.

Backtest on history the model has not seen. Train the model on data up to, say, the end of last year, and then let it forecast the months you already have the right answers for. Compare the errors of the model and the baseline, for example as the mean absolute percentage error, a metric known as MAPE. If the model does not win this comparison, it has no business being in your operations.

Retrain regularly. Demand changes, and a model that is trained once and forgotten quietly goes stale. A forecasting model is a process, not a project that gets finished: new data in, new forecast out, errors logged and compared against the baseline.

The data is closer than you think

The first question is always what data is needed, and the answer is usually: the data you already have. Sales history from the accounting system, till reports from POS systems and web traffic from analytics tools are the core. Two to three years of weekly figures are enough to capture both the trend and the annual rhythm.

To this you can add open Icelandic data. Statistics Iceland publishes consumption and price time series on px.hagstofa.is that give context on overall demand and price trends. The Icelandic Met Office offers open weather data that can be used directly as an explanatory variable where weather moves sales. The public tourism dashboards show tourist arrivals and overnight stays, a key variable for every company that feels the flow of tourists.

Where does forecasting pay off, and where does it fail?

Four areas of use stand out, along with the caveats that come with them:

• Inventory and purchasing. Forecasting demand by product category makes it possible to order ahead of the season instead of chasing it. The benefit shows up as less shrinkage and fewer empty shelves, both of them measurable.
• Staffing and shift planning. Forecasting traffic by day of the week and time of year moves the shift plan from gut feeling to numbers. The same goes for the timing of campaigns and the distribution of marketing budget over the year: put the effort where demand will be, not where it was.
• Uncertainty intervals matter more than point forecasts. A forecast that says "850 units" is less useful than one that says "between 700 and 1,000 with 80% probability". Decisions about inventory and staffing are about coping with the range, not about hitting a single number.
• Forecasts fail under structural change. A pandemic, a new competitor or a changed product range breaks the pattern the model learned from history. Then human judgement has to take over until new history accumulates. And a simple model you understand and can challenge is worth more than a complex model nobody in the company can explain.