London's Divide Was Called Character. It Was Actually Policy.

I built a machine learning model to find London's divide and you can enter your postcode to see which side you're on. We've been blaming the wrong people for it.

I have a bit of a weird habit. I cycle to a random café in a random London borough, sit down, and listen. I’ve done it for long enough to notice an invisible line: while cycling, at some point the shopfronts change, the language changes, the streets narrow, then widen. The café itself tells you almost nothing. Good coffee is everywhere now. What changes is everything around it: who’s in it, and what they’re talking about. (I now also know far too much about Terry, Steve, Brenda, and Olivia - their lives, their grievances, their strong opinions about the council and the weather.)

I’ve written about London’s pub closures, its restaurant algorithm, its segregation patterns. The whole time I’ve been circling the same question from different angles: is London’s divide just a feeling everyone recognises, or is it stable and structural enough, that you can actually draw it?

So I tried. I built a model using neighbourhood-level data across London: education, occupation, ethnicity, house prices, cafés, pubs, restaurants, ratings, etc. and asked a simple question: if there really is a boundary in this city, can a machine find it?

It could (and you can check where your favourite neighbourhood sits at the end). And once it did, the more interesting question became who put it there. The answer is less about the market than we've been told, and more about three policy decisions that are still sitting on the tab.

---

Sydney named it. London aestheticised it.

I was just a few months old when I landed in Sydney for the first time. What I didn't notice then was that Sydney was sharply and structurally divided. Sydney has named its divide: the latte line. The line between those who make the latte and those who drink it. Once you name a line, you have to argue about it. You have to explain who drew it, who benefits, and who’s been paying for it. Sydney has been doing exactly that, in newspapers, planning documents, parliamentary inquiries, for twenty years.

London has the same divide. Charles Booth, a Victorian social cartographer who walked every street in London and drew what he found mapped it in 1889. Wealth clustering in the west, deprivation running east along the Thames. The gradient he found is still visible in today’s data and has barely shifted in 130 years.

But London never named it. Instead, London called it character. Called it up and coming. Called it regeneration. A vocabulary carefully chosen to describe a structural inequality as if it were an aesthetic phase, something a neighbourhood passes through on its way to becoming somewhere worth living. We have been looking at this line for longer than Australia has existed as a federation.

So I wanted to do the impolite thing and draw it.

---

How do you find a line nobody drew?

The obvious way is to stare at a deprivation map until a story presents itself. I did something only slightly more dignified.

I assembled a neighbourhood-level dataset from all data in my previous Substack posts (google map’s algorithm, pub closures, segregation and inequality measurement) at Lower Super Output Area level of roughly 1,500 residents each, combining measures such as education, occupation, ethnic composition, house prices, café density, pub density, restaurant density, reviews and ratings. In other words, both the deep structure of the city, and the softer cultural signals people usually use to talk about class without saying class.

With thirty-odd correlated variables, the first task was compression. This is where Principal Component Analysis comes in. Highly educated areas tend to have higher house prices, more managers, more coffee shops, PCA notices that and summarises it. The result is one clean score per neighbourhood: essentially a poshness gradient. Once every neighbourhood has a poshness score, K-Means asks a deceptively simple question: if you had to split London into exactly two groups, where would the cut go? The result is a binary label for every neighbourhood. No geography was used. The final step was validation. A gradient boosting classifier which is basically an ensemble of shallow decision trees, each one learning only from the mistakes of the last, was trained to predict which side of the divide an LSOA falls on. It achieved 95% accuracy.

The point was to let the algorithm tell me where the line was, not to pick a line and ask the data to flatter my intuition. If the model could reliably classify areas as above or below the divide, a real boundary existed. If it couldn’t, the divide was a story we were telling ourselves.

The model found a line. Unfortunately.

Broadly, the line does what you’d expect. West is more affluent, east more working class, with the Thames doing its usual work as both river and sorting device. It catches parts of outer west London that the usual west-equals-wealthy story ignores. It puts parts of inner east London on the prosperous side earlier than the gentrification discourse usually starts paying attention.

Then I looked at what the model had actually used to draw it.

---

Beyond the usual suspects

I looked at what the model had actually used to draw the line. The café variables I'd spent considerable time scraping and quietly feeling proud of such as pub density, restaurant reviews, average menu price, were sitting near the floor. The model did not care about the flat white.

Then I looked at the top.

And there it was. Median house price: dominant to the point of embarrassment. Of course it was. House prices are where cities compress decades of decisions into a single number. They absorb transport access, school quality, planning restrictions, public investment, demographic sorting, speculative belief and inherited advantage.

But that’s exactly why house price is analytically useless as an explanation. It is an outcome masquerading as a cause. Asking the model to explain London’s divide using house prices is like asking why some people are rich and being told: because they have more money. Technically true. Analytically useless.

So I stripped them out.

What emerged in the house-price-free version was that now no qualifications jumps to the top. Ethnicity follows. The other factors that house prices were summarising all along become the lead driver the moment you stop letting the property market absorb the signal.

I checked this with SHAP values (my favourite method from cooperative game theory), which confirmed that the pattern was not being driven by a few outlier neighbourhoods.

---

The line wasn’t natural. It was built.

The model found the line. What draws it: house prices first, then credential sorting. That’s the model’s answer. But a model that tells you where the line is, and even what marks it, still hasn’t told you why it’s there.

Three decisions in particular turned that old geography into something durable. It all started with the wind though. London’s wealthy merchants did not choose the west out of aesthetic preference. The prevailing westerlies carried smoke from the City eastward, so they built upwind. That is not a metaphor. It is literally where the wind blew. And it created a geography that 400 years of policy have since worked very hard to preserve. That history is less elegant than a wind direction. It is also more hopeful, because policy can be changed.

The 1947 green belt froze London’s spatial structure just as postwar migration was reshaping who needed housing and where. Right to Buy, from 1980 onward, transferred council housing wealth to its beneficiaries, but the sold stock was never replaced. Then, in 1981, the London Docklands Development Corporation stripped democratic planning oversight from a large section of the East End and handed it to private developers. The DLR was built for Canary Wharf. The Jubilee line extension was built for Canary Wharf. Tower Hamlets, which surrounds Canary Wharf on three sides, was left with some of the highest child poverty rates in the country.

The map above shows what happened next. The working-class share of London’s population fell almost everywhere between 2001 and 2021: down 5.5 percentage points inside the Latte Line, and 8.6 outside it. But the revealing part is not who left. It is who replaced them. Inside the line, departing working-class residents were replaced by professionals; investment followed, schools improved, and the streets got the coffee shops. Outside it, the working class shrank just as fast, and fastest of all in neighbourhoods 2 to 4 kilometres from the Canary Wharf towers. Those areas absorbed the displacement pressure of a financial sector that needed nearby land and nearby labour, without receiving much of the investment that sector generated.

That is the uncomfortable corollary of blaming policy: the divide is not natural, inevitable, or the residue of some neutral market process. It was built. Which is, depending on your disposition, either the most depressing or the most encouraging thing about it.

---

Send the invoice

Three decisions helped build the divide wall. Three decisions, or their reversal, could begin to dismantle it.

First, undo what the green belt locked in. The 1947 Act froze a geography of affordability that hardened into a geography of deprivation. The answer is not to abolish the green belt, but to build aggressively within and around the boroughs it stranded. Social housing in Newham, Barking, Tower Hamlets and Southwark should be treated as infrastructure, not as a residual category for people the private market does not want.

Second, replace what Right to Buy sold and never replaced. The policy transferred housing wealth to its beneficiaries, but the stock itself was allowed to disappear. The result is an ever-smaller supply of social housing concentrated in exactly the places private capital has been least willing to go. One-for-one replacement - legislated, funded, enforced - is the minimum.

Third, send Canary Wharf the bill. The Docklands experiment was allowed to bypass democratic planning on the promise that its gains would spill outward. The data suggests something closer to the opposite: surrounding neighbourhoods absorbed the displacement without sharing in the uplift. In the areas just beyond the towers, the working-class share fell by 8.6 percentage points in twenty years. That figure should be in front of a Select Committee. A retrospective community benefit levy on Canary Wharf landholders, ringfenced for Tower Hamlets and inner Newham, would be politically controversial. It would also be one of the most analytically defensible things City Hall could do.

---

Which side are you on?

I can now place exactly where Terry, Steve, Brenda, and Olivia from the cafés I visited stand. I'll probably never see them again though (although I’m lowkey curious if Steve is still seeing that girl).

Enter your postcode at laurenleek.eu/latte_line_map and it will tell you which side your coffee place sits on, how it scores, and what variables are doing most of the work.

Despite my urge to try a coffeeshop in every London neighbourhood, I’m leaving London in a few days. We are cycling parts of Australia's east coast, starting in Sydney, where I intend to test the latte line empirically, from both sides. (I've been warned the flat whites can beat London. I'll report back.)

But Sydney isn't just a coffee benchmark. It named its divide. Naming it like Sydney is all this piece is trying to do for London. Not solve it. Not flatten its complexity into one model. Not pretend a gradient boosting classifier can explain everything that Charles Booth, planning law, migration, austerity, and capital have conspired to produce over four centuries.

Just name it. So we have to argue about it.

---

One last thing: I’ve just turned on paid subscriptions. Everything stays the same though. I’m still aiming to publish at least once a month, but if you’ve enjoyed this kind of work and want to help keep it going, consider subscribing or buying me a coffee. It means more maps like this one, more of the datasets behind them, and more of the ideas I’m genuinely excited to build next.