Data Science
Demand outlooks, sales forecasts and scenario planning built on your own history: the number, the range to plan around, and what changes if your assumptions do. Where the question comes back every month, the forecast reruns on its own.
Sales, demand and volume plans all rest on a forecast. We build it from your history, test it against simple alternatives on periods it hasn't seen, and give you the range and the assumptions with the number.
Forecasting demonstration · Stats SA retail trade sales
Forecasting
The business problem
Most organisations can't say how accurate their forecast has been, or whether a more complex method would do better than repeating last year. A plan built on an untested forecast carries a risk nobody has measured.
What the data shows
For South African retail trade sales, the tested forecast for Aug 2026 to Jul 2027 totals about R1,232.6 bn at constant 2019 prices, level with the last 12 months (R1,232.6 bn), because the winning method repeats last year's pattern. Neither statistical model beat it. Any growth the plan assumes on top is an explicit choice, and the scenarios below show what it does to the answer.
| Month | Forecast | 80% range | 95% range |
|---|---|---|---|
| Aug 2026 | R96.0 bn | R90.0 bn to R102.0 bn | R86.8 bn to R105.2 bn |
| Sep 2026 | R95.9 bn | R89.9 bn to R101.9 bn | R86.8 bn to R105.1 bn |
| Oct 2026 | R98.1 bn | R92.1 bn to R104.1 bn | R88.9 bn to R107.3 bn |
| Nov 2026 | R117.9 bn | R111.9 bn to R123.9 bn | R108.7 bn to R127.1 bn |
| Dec 2026 | R138.8 bn | R132.8 bn to R144.8 bn | R129.7 bn to R148.0 bn |
| Jan 2027 | R97.3 bn | R91.3 bn to R103.3 bn | R88.1 bn to R106.5 bn |
| Feb 2027 | R95.2 bn | R89.2 bn to R101.2 bn | R86.0 bn to R104.4 bn |
| Mar 2027 | R99.4 bn | R93.4 bn to R105.4 bn | R90.2 bn to R108.6 bn |
| Apr 2027 | R96.7 bn | R90.7 bn to R102.7 bn | R87.5 bn to R105.9 bn |
| May 2027 | R100.9 bn | R94.9 bn to R106.9 bn | R91.7 bn to R110.1 bn |
| Jun 2027 | R96.8 bn | R90.8 bn to R102.8 bn | R87.7 bn to R106.0 bn |
| Jul 2027 | R99.4 bn | R93.4 bn to R105.4 bn | R90.2 bn to R108.6 bn |
R billion at constant 2019 prices. SA Informatics analysis of Stats SA retail trade sales.
The decision it informs
The central path and range for a 12-month sales or demand plan, and whether a more complex forecasting method is worth paying for.
To act on it, you'd also need the drivers your plan depends on, such as prices, promotions and store changes, if a driver-based model is to be tested against the baseline.
What you would receive
A forecast with its ranges and assumptions, an evaluation record showing how the method was chosen and how it performed on unseen periods, and a process that reruns on each new month of data.
Why the question recurs
Plans are revised monthly or quarterly. Each new month of data is a chance to re-forecast, and to check the last forecast against what actually happened.
Questions like this
Data and method
Recommend a model only if it beats last year's pattern on months it hasn't seen, and its ranges hold at least 80% of outcomes. Otherwise, recommend last year's pattern. Every method is scored against the same yardstick: how its error compares with simply repeating last year.
A forecast is only as useful as what you know about it: which method to trust, how far ahead, how wide the real uncertainty is, and what the outlook assumes. Each tab gives one answer from the same evaluation, and what it means for a plan.
Forecasting demonstration · Stats SA retail trade sales
A full evaluation on Stats SA's published retail trade sales: the process we would run on your own sales history.
Each method forecast 12 months ahead from 16 points in time, 192forecasts each, scored only on months it had not seen. Repeating last year's pattern had an error score of 0.758 (lower is better). Exponential smoothing scored 0.767 and seasonal ARIMA 0.788.
The three finished too close to call, so under the rule set before the results were in, the simpler method is the recommendation.
What it means for a plan
Use the simple method's path as the central estimate. On this evidence, a more complex model does not buy accuracy.
Forecasts made in July 2025 against what happened next
The most recent of the sixteen tests. R billion at constant 2019 prices.
| Method | Error score (lower is better) | Forecasts scored |
|---|---|---|
| Last year's pattern | 0.758 | 192 |
| Exponential smoothing | 0.767 | 192 |
| Seasonal ARIMA | 0.788 | 192 |
One month ahead, seasonal ARIMA was clearly more accurate, with an error score of 0.47 against 0.67for last year's pattern.
Beyond the second month, no method was consistently ahead. The order changes from one horizon to the next.
What it means for a plan
If the decision only needs next month's figure, a horizon-specific evaluation of seasonal ARIMA may be worth running.
Accuracy by how far ahead the forecast reaches
Error score by months ahead, averaged over 16 tests (lower is better).
| Months ahead | Last year's pattern | Exponential smoothing | Seasonal ARIMA |
|---|---|---|---|
| 1 | 0.67 | 0.69 | 0.47 |
| 2 | 0.64 | 0.55 | 0.61 |
| 3 | 1.01 | 0.81 | 1.17 |
| 4 | 0.70 | 0.91 | 0.73 |
| 5 | 0.63 | 0.65 | 0.64 |
| 6 | 0.96 | 0.85 | 1.35 |
| 7 | 0.72 | 0.89 | 0.85 |
| 8 | 0.64 | 0.63 | 0.65 |
| 9 | 0.91 | 0.82 | 0.89 |
| 10 | 0.75 | 0.91 | 0.76 |
| 11 | 0.62 | 0.67 | 0.56 |
| 12 | 0.84 | 0.82 | 0.76 |
A range is only useful if it holds. Ranges meant to hold the outcome 80% of the time held it 96% of the time for the recommended method. The 95% ranges held it 100% of the time.
The 2020 lockdown inflates the error the ranges are built from. Checking this is part of every evaluation we run: most forecasts are never checked against the ranges they claimed.
What it means for a plan
Treat the ranges as cautious bounds, not precise ones. Planning on them errs on the side of caution.
Did the ranges hold?
Share of actual values inside each method's ranges, across 192 test forecasts.
| Method | Inside 80% range | Inside 95% range |
|---|---|---|
| Last year's pattern | 95.8% | 100.0% |
| Exponential smoothing | 95.3% | 98.4% |
| Seasonal ARIMA | 98.4% | 99.0% |
Repeating last year implies a flat year by construction: a 50% chance of coming in lower. Exponential smoothing puts that chance at 46%, and Seasonal ARIMA, which carries more trend, at 26%.
Sensitivity: if a shift the model can't see moved sales 1% lower, the chance of a lower year under the recommended method rises from 50% to 78%.
What it means for a plan
If the plan assumes growth, record it as an assumption, not a finding, and show how much it moves the answer.
How the 12-month outlook depends on the method and the history assumed
Median and 10% to 90% range of the 12-month total, R billion.
| Method · history | Median total | Chance of a lower year |
|---|---|---|
| Last year's pattern · full history | R1,232.5 bn | 50% |
| Last year's pattern · last 8 years only | R1,232.4 bn | 50% |
| Last year's pattern · last 5 years only | R1,232.5 bn | 50% |
| Exponential smoothing · full history | R1,237.5 bn | 46% |
| Exponential smoothing · last 8 years only | R1,238.1 bn | 45% |
| Exponential smoothing · last 5 years only | R1,242.9 bn | 29% |
| Seasonal ARIMA · full history | R1,261.9 bn | 26% |
| Seasonal ARIMA · last 8 years only | R1,262.0 bn | 32% |
| Seasonal ARIMA · last 5 years only | R1,257.9 bn | 11% |
Which choice will the estimate inform, and how good must it be to change it?
In the demonstrationThe 12-month sales plan: a central path, a range, and whether a model beats repeating last year.
Fix how methods will be compared, and the simple baseline they must beat, before any results are in.
In the demonstrationThe question, series and selection rule were committed before any values were analysed.
Score every method on periods it was not built on, the way it will be used in practice.
In the demonstration16 rolling forecasts; 192 scored forecasts per method.
A complex model earns its place only by doing better. Simpler methods are cheaper to run and easier to trust.
In the demonstrationNeither model beat the baseline, so the baseline is recommended.
Every forecast comes with a range to plan around, checked against what actually happened.
In the demonstrationThe 95% ranges held 100% of actual outcomes.
Where the question recurs, build the method to rerun on new data and reproduce exactly.
In the demonstrationPinned to one official release: a rerun reproduces every result exactly.
Forecasting demonstration · Stats SA retail trade sales
Every step from question to reproducibility, pinned to one official release, with the selection rule written down before the results.
Within data science, machine learning and language models are methods we use where a question needs them. Where the goal is to take repeat work off a team, such as turning enquiries into quotations, that's Applied AI: an application built on your knowledge, rules and systems.
A model that answers a recurring question should rerun on new data without being rebuilt each time. Where that is worth doing, we build the data foundation and the method into a system your team keeps using.
Describe the question, the data you have, and whether you'll need the answer again. We'll tell you whether we're the right fit.
