Created by AI, Curated by Experts: The Loculyze Way of Analysing Retail Locations

Why We Choose Accuracy Over Speed

Our team has over a decade's experience analysing mobile signal data, and we've learnt the hard way about the many pitfalls of this dataset and its wild spatial and temporal sample fluctuations.

We are often asked about mobile-signal analytics platforms that deliver glossy insights at rapid speed and low cost.

You may have seen one. Draw a polygon, hit “run”, and a few seconds later you get a neat set of charts. Under the hood, the process is usually the same: split the world into cells, count devices in each cell, apply a weight then add up the cells that intersect with whatever shape the user has drawn.

It feels precise. It’s really just counting raw pings. And if there is one rule of mobile signal analysis it is that you cannot base analysis on raw pings.

Fast? Absolutely. Accurate enough to support multi-million-dollar retail decisions? Not even close.

Below is how we do it, and why we’re comfortable being the opposite of “instant”.

1. We start with detection likelihood

Mobile-signal data doesn’t see all customers equally.

• Enclosed centres with long dwell times and high visit frequency get rich samples.

• Open-air main streets with fast flows and short stops get thin samples.

The typical self-serve platform treats those the same: draw your line, aggregate the H3 cells inside, and that’s your “traffic”.

Do that and you’ll think the enclosed centre is “high performing” and the strip is “quiet”, even when published metrics tell you the opposite. You’re measuring detection, not customers.

Loculyze trains a dedicated machine learning model for each destination, refreshed monthly, to correct for structural and behavioural differences in detection likelihood.

We don’t assume every ping is created equal. We model how likely it was to appear in the first place.

2. Time windows need statistical footing

Short time windows look sharp on a dashboard. “Last Thursday’s lunch surge” sounds precise. Often it’s just noise.

If your platform happily serves “last week vs this week” on every location, it’s usually because it isn’t asking whether there’s enough sample to make that comparison meaningful.

We flip the process:

1. Start with a Moving Annual Total (MAT) to get long-run stability.

2. Then drill down to monthly, weekly, daily and hourly views, with the MAT as the anchor.

That way you don’t mistake a temporary blip for growth, or panic over a “drop” that’s just sampling error dressed up in a bar chart.

3. We model journeys, not isolated pings

Detecting a customer in a centre is straightforward: they were there, and we saw their device.

Understanding where else they went is harder. The chance of detecting that same device again at the centre next door is much smaller. If you just “look for pings” in two polygons and call that cross-visitation, you will massively understate true behaviour and misread cannibalisation and synergy.

The fast way:

• Count devices in polygon A, count devices in polygon B, intersect the two lists, call that “shared customers”.

The Loculyze way:

• Build a four-hour journey window for each unique detection trail.

• Apply probabilistic cross-visitation modelling to understand the probability of detecting a ping along this journey.

You get a realistic view of trade areas and journeys, not a coincidence map of where phones happened to be caught twice.

4. Built for confidence, not for speed

You can’t do any of the above properly in a few seconds.

• Modelling detection likelihoods per destination

• Reconstructing journeys

• Normalising time windows

• Validating model outputs against ground truth

All of that takes compute, iteration and review.

So the Loculyze platform does not claim to be “instant”. We take a business day or two to turn around a run because we’d rather give you something you can take to an investment committee than something that just renders fast in a browser.

Most self-serve platforms are optimised for engagement: draw shape, see chart, feel smart. They’re great for curiosity and quick directional questions. They’re a poor foundation for buying a centre, reshaping a network or underwriting a redevelopment.

With Loculyze, you trade a little time for a lot of confidence.

5. Sample transparency, always

You’ve probably heard the line: “mobile data captures about 70% of the population.”

That coverage can happen—for the very largest locations, over long periods, when enough signals accumulate to smooth out detection issues.

Most of the time, it’s lower and it varies heavily by place and time-period.

Many platforms hide that behind a single “visits” number and hope you don’t ask too many questions.

We go the other way. Every Loculyze module shows:

• Unique customers sampled

• Total visits sampled

right next to the modelled outputs.

If you can’t see the underlying sample, you can’t judge reliability. And if we didn’t show it, we wouldn’t expect you to trust the numbers.

6. Prescriptive, not descriptive: the LoculRetail score

We’re not trying to drown you in charts and hope you find something interesting.

Our focus is squarely on Loculyze Sales Index:

• We combine visitation, engagement and customer value.

• We turn that into a single dollar-based index of expected sales.

• You use that index to compare and prioritise locations on one scale.

Well-managed centres should outperform the index. Under-managed or poorly positioned assets should lag it. Either way, you’re working from one clear, consistently applied valuation signal, not juggling six different metrics.

This is “created by AI, curated by experts” in practice: the models do the heavy lifting; our retail and analytics team make sure the outputs are simple, usable and aligned with how real decisions are made.

Proof: why aggregation isn’t enough

Here’s the uncomfortable truth for “draw a polygon and aggregate H3 cells” platforms:

If mobile signal data was reliable enough to use raw, you wouldn’t see huge swings in coverage by location, time and behaviour.

But you do. That’s why it has to be modelled, not just counted and up-weighted.

To build our model we trained mobile signal data against published traffic and turnover data across hundreds of centres and landlords. The question was simple:

Can a properly modelled view of visits and engagement match what the real world is reporting?

Loculyze Visits vs Reported Traffic

Across that portfolio:

• Correlation (Pearson r): 0.94, with the 95% confidence interval between 0.90 and 0.96.

  • Real-world traffic swings are mirrored almost perfectly by our model.

• R²: 0.95.

  • 95% of the variation in reported traffic is explained by our visits model.

• Loculyze / reported: 0.985.

  • On average we under-predict traffic by about 1.5%.

That under-prediction is deliberate. We deduplicate visitors where traditional traffic counters can’t. Counters see entries. We see people.

You do not get numbers like that by just counting pings in H3 cells and multiplying by a constant “coverage factor”. The sample is too uneven. Without modelling, the errors compound.

Visits as a sales predictor

Even before we bring tenancy mix or segment-level spend into the model, a single deduplicated visit:

• Explains around 60% of the swing in annual turnover.

That’s a strong base signal for understanding value—again, only possible when you model detection properly instead of treating raw device counts as ground truth.

Loculyze Estimated Sales vs Reported Turnover

When we take the final step and estimate sales using our full model:

• Correlation (Pearson r): 0.96.

  • Estimated sales and reported turnover move almost in lockstep.

• R²: 0.95.

  • 95% of the variation in reported sales is explained by our model.

• Loculyze / reported: 0.97.

  • We under-predict turnover by about 3% on average.

Our sales index tracks real-world performance closely but remains conservative. Well-run centres beat it. Under-performers don’t.

Again: you don’t get this kind of alignment by stacking H3 cells and waving an up-weighting factor at them. You only get it by modelling detection likelihood, journeys and time windows explicitly.

Why this matters in a world of instant analytics

Instant, self-serve platforms are built for speed and scale. They give everyone a dashboard and let them draw shapes.

That’s fine for marketing teams looking for broad patterns or for answering “what’s happening roughly over here?”.

It’s not fine when the question is:

• Should we buy this centre?

• Where do we put the next store?

• Which assets deserve capex and which don’t?

Location performance is driven by:

• Detection likelihood

• Visit frequency and dwell

• Journey structure

• Time-window stability

• Sample size and bias

If the system isn’t explicitly modelling those, you’re getting fast answers based on incomplete context. It’s convenience analytics: good for a quick look, dangerous as a decision engine.

Loculyze is built on the opposite philosophy:

Take the time to model reality properly, then give the client a clear, prescriptive answer—created by AI, curated by experts.

So yes, we take a day or two. In exchange, you get:

• A per-destination model of detection likelihood.

• Noise-resistant time series anchored on MAT.

• Probabilistic journeys rather than coincidental pings.

• Full sample transparency.

• A single, comparative sales index you can actually act on.

If you’re choosing where to put your next store, which centre to buy, or how to reposition a portfolio, that trade—confidence over instant—is the only one that makes sense.