How predictive analytics can tell you which customers are about to leave your venue… before they do
There is a customer right now — in your database, from last Tuesday — who is about to leave and never come back.
They ordered once. Enjoyed it. Meant to return. But they haven't. And without the right system in place, you won't know they're gone until it's too late to do anything about it.
Traditional hospitality analytics is almost entirely backwards-looking. You open your dashboard and see what happened: how many covers, what the average spend was, how many repeat orders came through. Useful information. But it answers the wrong question. The question that matters isn't what happened — it's what's about to happen. Which customers are drifting? Which ones are on the fence? Which ones are three interactions away from being lost?
This is the problem Rydra's predictive analytics engine is designed to solve. And understanding how it works — at least at the conceptual level — will change how you think about your customer base entirely.
Every customer is in a state
The starting point is deceptively simple. Rather than thinking about your customers as a mass of transactions, think of them as being in one of a small number of behavioural states at any given moment.
A customer is either:
Browsing — they've seen your venue, they're looking, but they haven't committed
Viewing an offer — they've engaged with something specific, a deal, a menu item, an incentive
In cart — they've shown real intent; something is sitting in their basket
Ordered — they've converted; you have their money and their data
Returned — they've come back; this is where loyalty begins to crystallise
Churned — they've gone quiet; they've exited the active relationship
These aren't arbitrary categories. They map directly to observable behaviour — clicks, sessions, orders, time elapsed — and every interaction a customer has moves them from one state to another. Browsing becomes deal viewing. Deal viewing becomes cart. Cart becomes order, or it becomes churn.
The critical insight is that where a customer is right now is the single most important predictor of where they go next. A customer in cart behaves differently from a customer who is browsing. A customer who has ordered once behaves differently from one who has never engaged. Their current state — not their entire history, not their demographic, not their preferences — is what determines probability.
Transitions, counts, and the emergence of pattern
Once you accept that customers move through states, the next question is: how do they move? How often does a cart become an order versus a dropout? How often does a first order lead to a return versus silence?
The answer comes from counting transitions. Every time a customer moves from one state to another, that movement is recorded. Over time, as thousands of customers move through the system, patterns emerge.
Not because people are simple. People are not simple. Individual behaviour is messy, emotional, and context-dependent. But large populations are patterned. This is the key insight that makes the entire framework work: while you cannot predict what one customer will do, you can predict with considerable accuracy what customers in a given state tend to do.
If you observe that 1,000 customers entered the browsing state, and:
400 moved on to view an offer
250 moved directly to cart
200 stayed browsing
150 exited
You can now say with confidence: a browsing customer has a 40% probability of viewing an offer next, a 25% probability of moving to cart, and a 15% probability of leaving.
That is not a historical observation. That is a forward-looking statement about the next customer who arrives in that state. You have moved from descriptive analytics — what happened — to predictive analytics: what is likely to happen next.
These probabilities, built up across all possible transitions between all states, form what the model calls a transition matrix. It is a complete map of how customers move through your venue's ecosystem — where they progress, where they stall, where they convert, and where they disappear.
Why churn is different from every other state
One of the most important design choices in Rydra's model is how it treats churn. Unlike browsing, cart, or ordering — states a customer can move in and out of — churn is treated as an absorbing state.
An absorbing state is one that, once entered, cannot be exited within the model's active window. The customer has left. The door has closed. There is no probabilistic pathway that naturally returns them to browsing or cart.
This might sound like a simplification, and in one sense it is — some customers do return after long absences. But treating churn as terminal is strategically essential, because it creates clarity and urgency.
Every transition either moves a customer toward continued engagement or toward exit. The system is no longer a neutral network of movements — it has an endpoint. And that endpoint changes how you interpret every state in the journey.
A browsing customer is not just someone who hasn't ordered yet. They are someone who is, at some probability, already drifting toward the exit. A cart customer who doesn't convert isn't just a missed sale. They're someone whose probability of returning is now lower than it was before they reached cart. Every moment of inaction has a cost, and the transition matrix quantifies exactly what that cost is.
This is how Rydra can tell you, for any customer at any stage, their churn probability — the likelihood that they will exit the system entirely within a specified number of future steps. Not what percentage of customers churned last month. What the probability is that this customer, right now, is about to leave.
For example, a typical model might produce outputs like:
Browsing customer, churn risk over next 3 interactions: 42%
Cart customer, churn risk over next 2 interactions: 22%
Post-first-order customer, churn risk over next 5 interactions: 31%
Returning customer, churn risk over next 5 interactions: 9%
These aren't abstract statistics. They're a forward-looking risk register for your customer base — updated in real time as customers interact with your platform.
The loyalty score: quantifying who is worth reaching
The inverse of churn probability is loyalty probability — and this is where the model produces its most commercially actionable output: the Rydra Loyalty Score.
The loyalty score answers a precise question: given where a customer currently is in their journey, what is the probability that they will become a returning customer before they churn?
It is not a measure of how much they've spent. It's not a frequency count. It's a forward-looking probability derived from their current behavioural state and the observed transition patterns of thousands of customers before them.
A typical output looks something like this:
Browsing: 0.28 — low engagement, significant churn risk
Viewed offer: 0.41 — responded to stimulus, meaningfully higher potential
Cart: 0.63 — strong intent, high intervention value
Post-first-order: 0.74 — on the threshold of becoming a loyal customer
These scores create a hierarchy of customer quality that doesn't exist in any conventional reporting tool. Your average spend dashboard treats a first-time orderer and a browsing user as categorically different only by whether they've transacted. The loyalty score tells you something far more nuanced: the first-time orderer is 2.6 times more likely to become a loyal customer than the browser — and both are worth pursuing, but in entirely different ways.
The marginal middle: where the real money is
Here is the counterintuitive insight at the heart of the entire framework — and the reason why Rydra's predictive analytics is built the way it is.
Discounts and incentives should not go to your most loyal customers.
Your highly loyal customers — those with loyalty scores of 0.8, 0.9, or higher — will return anyway. Offering them a discount doesn't change their behaviour. It just reduces your margin on a transaction that was going to happen regardless. You haven't created loyalty; you've simply subsidised it at your own expense.
The economically valuable targets are the customers in the marginal middle: loyalty scores in the 0.5–0.7 range. These are the customers who ordered once but haven't come back yet. The ones who got to cart and hesitated. The ones who viewed an offer three times without committing. They're not gone — but they're not secured either. They sit in a state of genuine uncertainty, and it is precisely that uncertainty that creates opportunity.
Because if a targeted intervention — a personalised offer, a well-timed reminder, a relevant deal — can move a cart customer's loyalty score from 0.63 to 0.74, that shift has measurable economic value. Not just one more order. A higher probability of becoming a repeat customer. A longer lifetime value. A compounding return on a single low-cost nudge.
The model formalises this completely. The value of an intervention is calculated as the incremental expected future revenue it generates, minus the cost of the incentive. If that number is positive, the intervention is worth making. If it's negative — if the customer was going to return anyway, or if the discount exceeds the incremental value created — it should be withheld.
This is the difference between promotional thinking and decision science applied to commerce.
The system gets smarter as it grows
There is one more property of this model that has significant implications for how Rydra thinks about scale: the system improves with every additional interaction.
Every order placed, every cart abandoned, every return visit, every dropout — each one refines the transition probabilities slightly. The model is not static. It is a continuously updating picture of how customers in your specific ecosystem, on your specific platform, actually behave.
This means that early on, the probabilities are approximations. As the platform grows — more venues, more users, more transactions — those probabilities converge toward accuracy. The law of large numbers applies: individual behaviour is noisy, but aggregate behaviour is structured, and at scale, that structure becomes increasingly reliable.
In practical terms, this means that growth doesn't just increase revenue. It increases intelligence. Every new customer who interacts with a Rydra-powered venue adds a data point that benefits every other venue on the platform. The predictive engine becomes more precise, the loyalty scores become more reliable, and the intervention recommendations become more accurate.
This is the strategic moat that makes behavioural analytics genuinely defensible over time. It is not a feature that a competitor can copy by building a similar interface. It is a compound asset that builds quietly through every transaction, every state transition, every customer journey recorded and learned from.
What this looks like in practice for venue operators
For a hospitality operator, the practical output of all of this looks nothing like a maths lecture. It looks like this:
"You have 47 customers who have placed one order and haven't returned in 18 days. Their average loyalty score is 0.61. Based on their current state, 23 of them have a churn probability above 35% within the next two weeks. These 23 are your intervention priority — here's what we recommend deploying, and here's the projected value uplift if they convert to returning customers."
That is actionable. That is the difference between a dashboard that tells you what happened and a system that tells you what to do about what's about to happen.
You don't need to understand transition matrices to use it. You need to understand one thing: your customers are always moving through a journey, their current position in that journey predicts their future behaviour with quantifiable probability, and the right intervention at the right moment — for the right customer — is not a guess. It is an informed decision grounded in data.
The summary
Most hospitality analytics is a rear-view mirror. It shows you where you've been.
Rydra's predictive model is a windscreen. It shows you where your customers are going — and gives you the tools to change the destination.
The system works by mapping customer behaviour into a small number of observable states, counting how customers move between those states across thousands of interactions, converting those movements into forward-looking probabilities, and surfacing two metrics that matter: churn risk and loyalty potential.
The customers who need your attention are not the ones already loyal, and not the ones already gone. They're the ones in between — sitting at cart, sitting at post-first-order, sitting at the inflection point where a single well-timed nudge is the difference between a loyal regular and a statistic.
That's the marginal middle. And that's precisely where Rydra looks.
Rydra is a hospo-tech platform connecting consumers to hospitality venues through a flat $1.99 platform service fee. No commissions. No margin extraction. Just owned customer data, predictive analytics, and the intelligence to act on it.
Want to see how the model works for your venue? Book a demo in the partner hub!
Sources & further reading: ClubRydra Behavioural Intelligence Framework (internal research, 2026) · Bain & Company: The Value of Keeping the Right Customers · BCG: First-Party Data and the Future of Personalisation · Olo: 60% of Restaurant Sales Are From Repeat Guests · SevenRooms 2024 Australian Restaurant Diner Trends Report