The number of people unemployed in the UK has fallen unexpectedly for the first time in 18 months… George Buckley [of Deutsche Bank] admitted previous predictions of the unemployment rate reaching 10% now looked unrealistic… [The figures] came as a surprise to many analysts.

But to regular readers of this blog, this news is anything but unexpected: in December 2009, my analysis of unemployment-related search trends clearly indicated that the unemployment rate was about to fall.

So could this be an example of search trends providing early insight into economic data? Possibly, but it’s only one month’s figures we’re talking about. A sustained track record of successful projection is needed to demonstrate that search analysis can yield valuable insights.

Over 2010 I’ll be keeping an eye on the data to see what happens. In the meantime, if you can think of other real-world metrics that might be suitable subjects for search trends analysis, get in touch.

Traditional economic reports, on the other hand, look at events that have taken place. Unemployment figures tell us how many people are claiming benefits rather than how many people have been put at risk of redundancy. Average house prices are based on completed transactions, not how many people are currently looking to buy. So while we can be fairly confident of these reports, they don’t provide us with particularly current insights.

This trade-off between confidence and currency was, in the past, largely academic as analysing current data was almost impossible. But in the age of the real-time web, this might be about to change: maybe patterns in search behaviour can give us a glimpse of future patterns in the economy.

We first became interested in this topic back in spring 2009, so we analysed search patterns for two sets of keywords as the UK economy went into recession. We looked for relationships between these search patterns and related economic indicators, and listed some tentative predictions based on what we observed.

House prices

In April 2009, we looked at volumes for 23 keywords that homebuyers might use, including buying a home, cheap mortgage and mortgage providers. UK search volumes for these keywords were then compared to house prices.

House prices charted against search volumes for 23 related keywords, from January 2004 to April 2009. Sources: Nationwide, Google Insights

Searches typically decline as autumn ends before rebounding in January. But in 2008, the January rebound was lacklustre and the decline came in spring – much earlier than usual. This was in line with house prices, which peaked in late 2007 and dropped severely from spring 2008.

In the first few months of 2009, however, search volumes enjoyed a far stronger January rebound than in the previous year – so we hypothesised that house prices would bottom out or even start to rise again in the middle of 2009. Let’s look at how accurate that hypothesis turned out to be.

House prices charted against search volumes for 23 related keywords, from January 2004 to December 2009 Sources: Nationwide, Google Insights

Sure enough, the search volume resurgence was accompanied by house price growth throughout 2009. But you’ll notice that search volumes soon tapered off, with a particularly steep fall after August. Our revised hypothesis, then, is that house prices will initially plateau and then drop again. We’ll revisit the statistics in spring 2010 to see how things turn out.

Financial difficulties

The second set of keywords we analysed was related to impending financial difficulties such as joblessness, debt and insolvency. They included signing on, mortgage arrears and debt problems, and were compared to the UK jobless rate.

UK unemployment rate charted against search volumes for 24 related keywords, from January 2004 to April 2009 Sources: Office for National Statistics, Google Insights

These search volumes dip at the end of each year before rising in January – and the rise in early 2008 was more pronounced than in previous years. The jobless rate started climbing three months later, suggesting that in this case search patterns might anticipate economic statistics. We observed that search volumes had dropped significantly in the first few months of 2009, so our hypothesis was that the jobless rate would stabilise but not drop between April and July. The chart below shows what actually happened.

UK unemployment rate charted against search volumes for 24 related keywords, from January 2004 to April 2009 Sources: Office for National Statistics, Google Insights

The unemployment rate has indeed stabilised, wavering between 7.7% and 7.8% since early June, suggesting that our original hypothesis was valid. And search volumes have kept on dropping throughout 2009. If search trends do anticipate economic reports in this case, we should see the unemployment rate drop steadily between now and spring 2010. Again, we’ll revisit these figures in April to see if this happens.

Conclusion

Our hypotheses from April 2009 were largely borne out as the year progressed: the drop in house prices was reversed and unemployment rates stabilised. So maybe there is some truth to the notion that search patterns can shed some light on forthcoming economic change.

But these hypotheses were in tune with the economic mood of the time. Many commentators were talking about green shoots and a V-shaped recession – there was a feeling that recovery was just around the corner. Today, we remain in what has become the longest-running recession in recorded history and there is considerable uncertainty about what 2010 will bring.

Our new hypotheses are less likely to be tainted by current economic consensus, precisely because no real consensus seems to exist right now. For this reason, the idea of search predicting the future will be seriously tested as the year unfolds. Don’t forget to come back in April 2010 to see the results for yourself.

Hitwise gathers its data by looking at the traffic logs of its clients’ websites, which number around 1,500. These websites are divided up into a range of categories and sub-categories. Hitwise is therefore aware of traffic volumes to sites in different categories, and has attempted to draw conclusions from changing patterns in these.

This methodology may be imperfect—1,500 websites may sound like a lot but is just a drop in the ocean—but the findings seem to be intuitively correct. For around a year now the subject of how a slowdown would affect online behaviour has been coming up more and more frequently, and the general consensus has been that online retail and price comparison sites are not going to be as exposed to the effects of a drop in consumer spending. Smaller household budgets lead to an increase in price-sensitivity, and price-sensitive consumers spend more time researching and planning purchases as opposed to buying on impulse.

An example of this in the Hitwise research is that traffic to what it identifies as price comparison, voucher or cashback sites increased by 20% between July 2007 and July 2008, after a slight drop in traffic to such sites between 2005 and 2007. Voucher sites seem to be the biggest beneficiaries (to the uninitiated, voucher sites collate promotional codes & vouchers from various retailers, which can be redeemed at checkout for discounts – here’s an example).

However, another contributor to this trend could also be quite simply that British people have become, on average, more sophisticated online shoppers. It’ll be interesting to look at how traffic to voucher and price comparison sites bears up when the growth phase of the next business cycle begins. Will those sites become the online equivalents of Poundstretcher, shunned by all but the most price-sensitive? Or will they remain the first port of call for the clued-up online shopper?

While putting this short-list together I’m visiting quite a lot of ‘showcase’ sites for each package on my long-list. And sheesh, some of them are bad.

I don’t mean “bad” in the sense of bad user experience design, even though it’s fair to say that many of them are guilty of that. I mean “bad” in that the products themselves are bad, some of them really bad.

It’s a consequence, I suppose, of the barriers to entry for e-commerce being so low these days. In fact, my preliminary exploration of open source e-commerce options has established that they’re even lower than I’d assumed them to be.

For example, I’ve come across an Australian site that sells t-shirts saying “The first rule about Kite Club is never talk about Kite Club”. Erk. It reminds me of a t-shirt I saw in Paris once, which I still think of as the worst t-shirt I’ve ever seen. It said, “the first rule about computer club is that you don’t talk about computer club”.

But aside from bad t-shirt slogans, of which there are plenty, the biggest culprits are the numerous arts’n'crafts retailers.

Before the internet, a lot of this stuff – the results of amateur pottery classes and the like – would have just been given to relatives or left to accumulate in cupboards and boxes. But now, the arguments against creating an online retail site for these efforts get weaker all the time as e-commerce gets easier. And it seems as though there’s a market out there for a lot of this twee, throwaway kind of stuff. That’s the “long tail” for you, I guess!

However critical I might sound in this post, though, I should point out that I’m not advocating the eradication of such sites from the internet. I’m just noting my vague fascination with this underbelly of online retail that I hadn’t really explored until today.

As the week of testing came to a close, it was possible to flick through the responses and get a sense of what adjectives were the most popular. However, it was less easy to convey this to the client in summary form.

Of the 100 options available, just over 30 had been chosen by at least one participant, meaning that rendering the results of the survey as a bar or pie chart would be at best inelegant and at worst unintelligible. And I couldn’t chop the least popular choices just to present a simple overview, as this would skew the data and paint an artificially positive picture of how the participants had responded to the site.

In the end I drew the results of the survey using a “word cloud” model. If you’ve used, well, the internet in the last couple of years you’ll have seen these (although the term itself may not be so familiar!). Each adjective that had been chosen at least once was displayed in the ‘cloud’, and its text size was determined by how many participants had chosen it. This meant that the most popular options stood out clearly and the less popular options, although less visible, were still legible if the diagram was studied closely.

The resulting cloud met with a positive reception when presented to the client and helped to provide a quick and effective summary of the test sessions, especially useful for people in senior management who didn’t have time to go through the detailed analysis of the tests.

Although I used Visio to create the cloud, there are a number of tools online that can be used to quickly generate word clouds of your own. Wordle, at http://www.wordle.com, is the one I’d most recommend.

I’m not convinced that they’re always useful but you never know when you’ll end up in a situation where a word cloud might come in handy.