< All Insights

How to predict your e-commerce traffic for the next year with just 7 numbers

Profile photo of Andrew Ladd

Andrew Ladd

Project manager

There's no shortage these days of articles about how unrecognisable the world feels in the age of coronavirus; we live, as almost anyone will tell you, in uncertain times.

In some ways, though, the times aren't as uncertain as they seem. Knowing what sort of business to expect on your website might be easier than you think.

The week lockdown started in the UK, we had the slowest week of traffic on the Ten4 website in recent history. (We looked back 18 months trying to find a slower one and then gave up.) Likewise, March was the slowest month we can remember seeing, with just 707 users visiting the site. No doubt many other websites — except the ones booking grocery delivery slots, naturally — saw similar slumps in traffic. It felt, at the time, apocalyptic.

But if you'd told us a year earlier that we'd be getting 707 website visits this March, we probably wouldn't have been worried, or even that surprised. Instead we would have said: sounds about right.

We’re not trying to claim we saw coronavirus coming. But we know enough about basic statistics to understand that all businesses ebb and flow over time. While coronavirus has had many real and terrible consequences so far, planning to get through this trying period means separating out those real consequences from the natural ebb and flow. So that's exactly what we did.

Use one week’s traffic to predict the rest of the year

So how did we know our traffic was going to slump in March? Well, the short and slightly disingenuous answer is that March is always quiet — the January rush of people thinking new year new website has passed, and entering the end of the financial year means budgets are already spent until April.

The longer and more honest answer is: we didn't really know March would slump, not exactly. But even a year ago we could have told you that the number of users this March would be somewhere between about 590 and 1,500, purely from random, week-to-week fluctuations. Put in that context, 707 is low, certainly, but it no longer feels so apocalyptic.

Of course, you might object that giving a range of nearly 900 potential outcomes isn't really predicting the future in any meaningful way. That's fair enough, and actually we have a far more detailed internal model that cuts down the range significantly.

What you should find impressive about the 590 to 1,500 figure is how we got there: it was calculated based on just 1 week of daily traffic from February 2019. That is, we took just 7 numbers and used them to correctly forecast our total monthly traffic every month for the following 15 months.

For that matter, the same 7 numbers let us accurately forecast the total weekly traffic for 58 out of the following 60 weeks. The week lockdown started was admittedly one of the only two weeks we got wrong, but we missed it by only 11 users. That's a margin of error we feel okay about when we're guessing from a year away.

Graph showing a close correlation between estimated website traffic and actual website traffic

The solid blue line shows our actual cumulative traffic over the year; the dotted grey line shows the prediction after just seven days of data.

Put your current traffic in context

The not-so-secret method behind these predictions is something called Monte Carlo modelling, which uses thousands of random datapoints to bulk out your real ones. Those 7 days in February were enough to extrapolate a pattern — a very rough pattern, to be sure, but a pattern nonetheless — so we generated about 14 years' worth of fake traffic data following the same pattern. We then analysed that fake data as if it were real — and, as it turns out, it might as well have been. That’s why we were able to get so many weeks right.

To some extent we got lucky. When you're extrapolating from just 7 days of data, your prediction is very sensitive to small variations; that's why the range, at 900, is so wide. If we'd picked a different week, it might have been wider still, or smaller, or higher, or lower. That's why our actual internal model uses a lot more than 7 data points and gets to a much more accurate number.

Still, there's definite value in even a wide range, because imagine if the number of visitors we'd had in March was only 200. To be that far outside even the very wide range would be a definite cause for panic, a sure sign that business was grinding to a halt and drastic measures were necessary.

But knowing that even the quickly plunging traffic in March was still within the realm of normality, we were able to pause and take a deep breath and hold our nerve. We kept going, albeit from home, as if business hadn't significantly changed. That was a good thing, too, because traffic for April, and then May, rebounded to normal levels, and business kept coming in, slowly but surely.

We got lucky in that respect too, and we know it, and there's of course no reliable way of knowing what the world might look like a few months from now. Armed with the numbers, though, we'll at least know whether our business is carrying on as usual — and whether we actually need to panic. And if the low months start to become more common than the high ones, those data points will feed into the model and we'll adjust our forecasts, so if nothing else we'll always know what direction we're heading in.

Graph showing the accuracy of data improving over time using Monte Carlo modelling

As the year goes on, the range of possibilities gets smaller and smaller, and by 6 months in the estimate for the year is within about 2% of the actual final figure — and this is all based on the original 7 days' data, without any adjustments as more data comes in.

Predict conversion rates and revenue too — and put them all together

In some respects the levels of traffic on our website are a bit irrelevant, because we don't actually sell anything there — our revenue is less directly tied to traffic than if we ran a full-blown e-commerce site.

But lots of people do run e-commerce sites — lots more people, since lockdown started — and for them this kind of forecast can be invaluable. The great thing about Monte Carlo models is that they can be used for just about anything, and they can be easily combined. So if you know the range of likely visitors to your site, the range of likely conversion rates, and the range of likely order values, you can come up with a pretty decent estimate of your range of likely revenue for any given period.

And while the month to month range may still be large, the other great thing about Monte Carlo models is that they get more and more accurate the further out you try to predict. The spread for our 7-day model's monthly predictions may have been 900, but over 14 months it missed the total number of visitors to our site by just 884 — less than one month's variance either way.

If you can predict your business's revenue to that level of accuracy for the next year, you're in a good position to decide whether to make cuts, investments, or neither. It can also give you a really clear picture of how small changes to your e-commerce business can affect your bottom line — for instance, if a digital advertising company promises to boost your traffic by 2%, you can easily see what this might translate to in revenue, and whether or not it's worth doing.

Even if you don't completely trust the final prediction, the extreme ends of the range will at least give you a good idea of the worst- and best-case scenarios you need to plan for. And in uncertain times, that little bit of certainty can feel like a real relief.

In the meantime, we hope you're staying safe out there.

Share

Profile photo of Andrew Ladd

Interested in learning more about predicting your website income? Talk to Andrew