£8,150.75
Best Single Month
22
Months In Profit
£101k+
Withdrawn
Picture the scene at a quarter past seven on an ordinary weekday morning.
Your phone buzzes on the bedside table with a new email.
It's a single message from Bet Compiler Pro, and it's short.
Inside it are three, maybe four selections, each one flagged by an AI model that worked through over 4.6 million data points overnight while you were asleep.
Each line gives you the horse, the course, the race time and the price worth taking.
You log into your betting account and the selections are on in under 8 minutes.
Then you close the laptop and get on with your day.
That's how the system works from your side of the screen, start to finish.
Usually by 7.15am
In under 8 minutes
£1,575.30 in the past 7 days
There's no form study involved, and you'll never again lose an evening fretting over each-way terms.
I know how that sounds, because I've read a hundred pages that open with a promise like it.
Most of them were written by people who'd never set foot behind a bookmaker's counter.
I spent eight years back there, settling the bets other people placed, and I can tell you the view from that side is nothing like the one you get from the shop floor.
You see which bets come off, which ones never do, and why the same people keep handing their money over the counter week after week.
That view is what shaped Bet Compiler Pro, and it's why the selections look nothing like the ones you'll find in any tipping column.
Let me tell you how Bet Compiler Pro came to be, and why I believe it's the most unfair advantage available to any punter in Britain right now.
I'm 51, I live in Worcestershire with my wife and four kids, and for the past two-and-a-half years I've been running the most reliable AI-driven horse racing selection system I've ever come across.
It matters partly because of the results, but mostly because of where I started.
I know what losing feels like, and I know the sinking feeling of watching a system you were sure of fall apart over six weeks.
I know what it's like to be up £1,200 in March and down £900 by April.
I spent nearly a decade going round in circles, chasing the next idea that was going to fix it.
Every winter I'd tell myself the new season would be the one where it all clicked, and every spring I'd be back where I began.
I'm writing this today because I eventually found a way out of that cycle for good.
It didn't come from studying harder or from paying another tipster.
It came from building something that works completely outside the way the average punter thinks.
Few of the people who use this service know this about me, but I spent the better part of eight years working in the back office of a regional bookmakers chain.
The head office sat above a carpet shop in Redditch, and the back office was a long room with no windows and a radiator that only had two settings.
The role was called a settler.
When a customer came in with a complex accumulator, a fold bet or a Lucky 63, it landed on my desk to work out and pay out.
In those days the slips came in by fax and by courier from 23 shops across the West Midlands, and each one had to be checked by hand before the money went back out.
I learned Rule 4 deductions, dead heats, each-way fractions and the knock-on effect of a non-runner in a Yankee before I'd been there a month.
I saw everything that passed through that office, from the biggest payouts to the long stream of losing returns.
I also saw the clever accumulators that nearly came off, and the dozens of little habits that kept ordinary punters losing.
I watched how they moved the market and how they reacted when money arrived from certain directions.
I watched how they used early morning data to set their tissue prices before racing even started.
A tissue price, if you've never come across the term, is the trader's first attempt at a betting market, drawn up before a penny has been staked.
Some mornings they'd price up 14 meetings before nine o'clock, working at a speed that would have made your head spin.
I'd watch them adjust a price three or four times before ten o'clock as the first bets came in, and I started to see how much of their opening market was educated guesswork.
The traders called it making a book, and the best of them could tell you within a pound or two how much each horse would take before the first race.
I absorbed all of that without fully realising what I was learning.
In 2016 the chain bought an automated settlement package from a software firm in Leeds.
Within 18 months the back office went from six settlers to one, and the one they kept wasn't me.
I remember walking out on my last afternoon with a cardboard box under my arm and a strange sense of relief.
The software could settle in a second what took me five minutes, and it never once miscounted a Heinz.
I didn't know it then, but that was my first real lesson in what a well-built program can do with a pile of racing data.
It took me another eight years to put that lesson to work for myself.
When I left and took up betting seriously as a punter, I was carrying eight years of inside knowledge about how bookmakers operate.
The trouble was that knowing how they worked didn't mean I could beat them.
I spent years searching, following tipsters who made strong starts and then faded away.
I built my own models at home that looked beautiful on paper and fell apart in the real world.
I went to seminars, read books and had long conversations on forums with people who swore they had it cracked.
One fellow on a forum sold me a "secret" staking plan for £97 that turned out to be a photocopied page from a 1980s betting book.
I'd have a solid three-month spell, then a rough patch that gave back half of what I'd made.
It was progress, just not the kind you can build a life around.
My records from those years show a small profit overall, but it was the kind of profit that vanishes the moment you divide it by the hours spent earning it.
There was a regular in one of our shops, a retired engineer called Len, who placed the same Yankee every Saturday for six years.
I settled hundreds of his slips, and I reckon he finished those six years about £4,000 down, despite being one of the cleverest men I've ever met.
By 2023 I was starting to suspect I'd become a younger version of Len, and that frightened me more than any losing run.
I'd tried more systems than I could count, and the only thing they had in common was that they worked brilliantly until I used them.
In early 2024 I was at a trade event in Birmingham, one of those industry gatherings where developers and analysts talk about machine learning applications.
I wasn't there for betting.
I was there because a friend had dragged me along to look at software solutions for a small property business he was running.
But I wandered into the wrong seminar room and ended up sitting through a 40-minute talk about predictive modelling in competitive sports.
The speaker was a data scientist called Ryan, in his late twenties, presenting to a room of maybe 15 people.
I found out later he'd studied mathematics at Warwick and spent three years building forecasting tools for a sports data company before going freelance.
Half of them were looking at their phones, and one man at the back was asleep.
I sat near the front with a free pen and a lanyard I'd forgotten to take off, and within five minutes I'd stopped thinking about my friend's property business altogether.
His point wasn't that people lack the brains for it.
It was that the relationships only show up when you line up thousands of results side by side, and nobody has the time or the memory to do that by hand.
He gave an example from football analytics.
A model had picked out six factors that the club's analysts had dismissed as trivial, and together they predicted the outcome of corner kicks at a rate 31% above random.
No human analyst had spotted that relationship, and it wasn't because anything was hidden.
To see it, you'd have to cross-reference six data streams at once, across thousands of corners.
That example stayed with me for the entire drive home along the M5.
Somewhere around Droitwich I realised that a corner kick and a horse race have something in common, as both are shaped by small factors that nobody bothers to add together.
The traders I'd worked alongside didn't rely on gut instinct.
They used morning data feeds, liability models and weighted probability tools to set their markets.
They weren't guessing, they were running calculations I couldn't see.
As a punter, I'd been bringing a notepad to a gunfight.
That evening, I dug out Ryan's details from the event programme and sent him an email.
I explained the eight years in the back office and the records I'd built up since.
I also told him what I wanted, which was a model that could look at a day's racing and tell me where the prices were wrong.
I expected a polite brush-off, but he rang me back within 48 hours.
He told me later that he'd almost deleted my email, because he got a message from a gambler every few weeks and none of them had ever had any data worth using.
What changed his mind was one line about 9 years of race-by-race records, kept the same way a settler keeps a ledger.
It had been designed initially for tennis, of all things, but the underlying architecture was sport-agnostic.
You feed it structured data, define the outcome you want to predict and set the learning parameters, and it goes looking for patterns.
More importantly, it tells you which patterns hold up over time and which ones are just coincidence.
Ryan explained that most betting models fail at that second job, because they find a pattern in last year's results and assume it'll repeat forever.
They covered every horse, course, going condition and class of race I'd followed.
They also held the trainer, the jockey, the stall position and the price movements for each runner I'd tracked.
Most of it sat in neatly categorised files, and some of it lived in old ledgers I'd scanned and typed up.
Ryan said it was the most thorough personal racing database he'd seen anyone keep outside a professional operation.
He also said it was a bit of a mess, which was fair, because I'd changed the way I recorded things at least four times over those 9 years.
Ryan agreed to take it on, and we settled terms over a call that lasted three hours.
It took about 6 weeks from that first call to having something worth testing.
The build phase took over every evening and most of my weekends.
During that time I fed Ryan every file I had, cleaned up the formatting to his instructions and learned more about feature engineering than I'd ever expected to.
Feature engineering means taking raw data and converting it into inputs a machine learning model can learn from.
It's technical work, but the principle behind it is easy to grasp.
Raw data tells you what happened, while feature engineering helps the model understand why it happened.
Ryan spent the first fortnight on the features alone, before he'd written a line of the model itself.
He was deciding how to present the data so the program could find real patterns instead of chasing coincidences.
That process alone was an education.
Take a horse's finishing position, for example.
Fourth of 16 by six lengths is one number on a form card, but Ryan turned it into several.
He worked out how far the horse was beaten relative to the winner's own rating, how much ground it made up from the two-furlong pole and whether the race was run at a crawl or a gallop.
He did the same with weight, converting raw pounds carried into a figure that accounted for the horse's age, the trip and the going.
Prices got the same treatment, because a 10/1 shot in a 6-runner race means something different from a 10/1 shot in a field of 22.
By the end of the fortnight, each runner in my 9 years of records had been rebuilt as a row of more than 60 inputs.
He went with something called gradient-boosted decision trees, which is a mouthful, but the idea behind it is clever and surprisingly intuitive.
Picture a long chain of yes-or-no questions about a runner, such as whether the ground has eased since its last run or whether its trainer has had a winner in the past fortnight.
One tree on its own makes a rough guess, and then the next tree is built purely to correct the mistakes of the first.
The third tree corrects what the second got wrong, and so on, until hundreds of them are working together and each one is a little more precise than the last.
At the end of the chain, the system gives every horse in a race a percentage chance of winning.
Those percentages have to add up to 100 across the field, which sounds obvious, but it was one of the first checks Ryan built in.
The tree-building runs on an ordinary cloud server that Ryan rents by the hour, and a full retraining on the complete database takes a little under 40 minutes.
The temptation with any model is to test it on the same races it learned from, and that tells you nothing.
He trained it on my first four years of races only, then asked it to make selections for the fifth year, which it had never seen.
Then he added that fifth year to its training, asked it to predict the sixth, and kept rolling forward like that through all 9 years.
It's the closest thing you can get to time travel with a betting system, because each prediction is made blind, just as it would be on a live morning.
The second check was something called calibration.
If the AI says a group of horses each have a 25% chance, roughly a quarter of them should win, not a tenth and not a half.
Ryan plotted those predictions against real outcomes on a chart, and the line ran close to the diagonal across almost every price band.
That was the moment I stopped thinking of this as an experiment and started thinking of it as a tool.
Most members never see this part, but it's worth knowing what goes on before the email lands.
Shortly after midnight, the system collects the final declarations for the day's racing, along with the overnight going reports and the early prices.
It then rebuilds the 60-plus inputs for each declared runner, which on a busy Saturday can mean well over 500 horses.
Each of those horses is run through the chain of trees, and every race ends up with a full set of winning chances that add up to 100.
Anything that fails the negative filter is struck out at this stage, however attractive its price looks.
What's left is ranked by how far the price strays from the AI's own figure, and only the strongest few make it into the email.
That's how 4.6 million data points turn into three or four lines of text on your phone.
By the time the first race goes off each day, the email has done its job and the system is already gathering data for tomorrow.
Horse racing generates an enormous amount of data.
There's every race, every trainer's season statistics and every jockey's performance across different track types and going conditions.
There's also every horse's full history, broken down by how it ran, where it sat in the race, how it finished and how its rivals performed next time out.
On top of that you have course configuration, class changes, weight shifts and the market's movements from tissue price to off.
Weather data, official going reports and going stick readings all come into it too, where they're available.
Ask even a seasoned form student to juggle 8 to 12 variables across a whole card and the quality of their judgement starts to slide.
60+
Inputs Processed For Every Race At Once, With No Fatigue
It handles them all at once, free from the distortion of having watched the horse last week and formed an opinion.
It also isn't swayed by the pressure of having told your mates to back something and then watching it drift.
It runs the same calculation every time, with no emotion and no favourites.
Where it really earns its keep is in what Ryan calls interaction effects.
That's where two, three or four variables combine to produce an outcome none of them would predict alone.
Now pair it with a trainer who has an outstanding record when moving horses up in class after a layoff, at a course whose layout suits its running style, on ground that's been independently tested as good-to-firm when the official going says good.
Suddenly that 20/1 shot looks very different.
A human tipster would have to hold all of that in their head at once, weigh it up and reach a conclusion, then repeat it for every runner in every race they cover.
Nobody can do that reliably, yet the model does it without breaking a sweat.
It runs every combination of factors in milliseconds, flags the selections where the evidence is strong enough and ignores the rest.
Most punters look at form and see a sequence of results such as won, fell, fourth, second, won.
They read that string of numbers and form an opinion.
It isn't interested in the headline results so much as what those results reveal about the horse's underlying performance.
Was it gaining or losing ground in the final furlong, and was it making up lengths or giving them away?
Did it travel well through the race, or did it waste energy fighting the rider before the turn?
Those details live in the race replay data that's fed into the system.
They paint a very different picture from the bare finishing position.
A horse that finished seventh but was the fastest thing in the race over the last furlong is a very different proposition from one that finished seventh because it ran out of puff.
On a form card they look identical, and the price usually treats them the same way.
Every horse has a profile of conditions it performs best in, covering the going, the course layout, the trip, the class of race and where it's drawn.
Experienced punters know most of these, but the AI goes further and studies how those conditions interact.
A horse that goes well on good ground might perform entirely differently on good-to-firm ground when a particular jockey is on board.
A course that favours a high draw in small fields might favour a low draw in fields of 12 or more.
Those small adjustments matter, and most humans, however experienced, aren't applying all of them to every runner in every race.
This is where my years in the back office become directly useful.
The system doesn't stop at starting prices, as it tracks each price movement from the morning tissue through to the off.
It's looking at the shape of each horse's price across the morning, not just where it starts and finishes.
One horse might shorten early and then ease before racing, while another sits untouched all week and then contracts in the final hour.
Those shapes aren't random, and my 9 years of records contain thousands of examples of each one, along with how those horses ran.
The AI has learned which price shapes tend to come before a strong run and which tend to come before a flop.
When the price shape lines up with the form profile and the conditions match, the model's conviction rises steeply.
That's where the best-priced selections come from, the 6/1 and 8/1 shots that the model rates as having the winning chance of a 2/1 or 3/1 horse.
Most systems ignore this part completely.
The system hunts for reasons not to back a horse just as hard as it hunts for reasons to back one.
Certain combinations of factors produce a repeated pattern of underperformance, and Bet Compiler Pro has learned to recognise them.
It'll exclude a horse that looks compelling to the human eye if the underlying data contains one of those warning patterns.
Trainers who struggle with certain jockey bookings get marked down, and so do horses showing a particular kind of flat finish in their recent runs.
All of it gets factored in, which is why there are fewer losers than you'd expect.
Finding winners is only half the job, and cutting out avoidable losers is the other half.
This layer has its roots in my settling days.
Layer 1
Form, read in depth
Layer 2
Conditions matching
Layer 3
Market intelligence
Layer 4
Negative feature filter
When people hear I worked for a bookmaker, they assume the useful knowledge was about how prices are set.
Some of it was, but the most valuable thing I took away was a mental catalogue of how ordinary punters lose.
After eight years of settling, I could pick up a slip and tell you whether it was likely to land before I'd looked at the results.
The same mistakes turned up week after week, from regulars who were nobody's fools.
They backed horses that had just won, at prices that already assumed they'd win again.
They doubled their stakes on the last race at the evening meeting to chase what they'd lost at lunchtime.
And they kept returning to the same handful of fashionable yards, whose runners were nearly always overbet.
When Ryan and I sat down to design the negative filter, I gave him a list of 40 of those habits from memory.
He turned each one into a pattern the model could test against the records, and most of them held up.
The AI now refuses to touch the kind of bet I spent eight years watching people lose on, and that alone saves members a small fortune.
It's a machine that finds the difference between the price on offer and a horse's true chance of winning, and bets only when that difference is wide enough to be worth backing.
It doesn't rely on luck or inside information.
The first version of the AI found winners, but it also found too many losers.
In the early weeks it was landing winners at decent prices, but too many of its losers were horses I'd have crossed off myself.
Ryan said that was to be expected, as the AI needed enough live outcomes to learn which of its early assumptions were wrong.
Each result went back into the system, so the patterns that predicted winners grew stronger and the weak ones faded.
It wasn't life-changing, but it was a positive start.
In week two, Ryan adjusted how the model weighted the going, and the profit that week came to £570.
The week after that produced 7 winners from 13 and a balance of £724.50 across the week.
I could tell something was working.
I had no plans to second-guess it, but I wanted to understand the logic.
What I noticed surprised me.
The AI wasn't picking horses that looked like obvious winners.
It was finding horses whose price was badly out of step with their real chance of winning.
That's the game in a sentence, because you find the horses the price has misjudged and bet when the misjudgement is big enough.
Professional traders spend their careers doing that by hand, and my model was doing it automatically each morning before I'd even got out of the shower.
Profit sat at just over £3,000 across my accounts, so I opened two new accounts and spread the risk.
Month two brought 35 winners from 53 and took the total banked past £7,200.
I started withdrawing slowly, keeping each balance at a level that wouldn't draw attention.
The major bookmakers run flagging software on their own customers, and an account showing a positive return over a 90-day window gets reviewed.
I knew how to keep my accounts looking unremarkable.
I spread the bets over several accounts, took regular small withdrawals and varied my stakes so no pattern could emerge.
When I wanted to get bets on at higher stakes, I'd walk into town and use the shops, picking a different one each time and paying cash so there was no account trail.
The system kept improving through the first three months before levelling off at a figure that has barely moved since.
Ryan told me at that point that the plateau was actually a good sign.
Models that keep climbing forever are usually overfitting, which means they're learning the coincidences along with the real patterns.
A model that finds its level and holds it is a model you can trust.
The first six months were all about learning where the model was making errors.
We learned which feature weights were out of line and where my own data had holes that needed filling.
Ryan sent me weekly summaries showing where the model had been most confident, cross-referenced against actual results.
Wherever it was most confident and most wrong, that's where we focused.
We rebuilt the going condition weighting three times in the first year.
After month four we added a stall-position correction, because I noticed the model was underperforming on straight-mile courses at certain tracks.
After month seven we brought in jockey booking data, because the pattern of jockey switches in the 48 hours before a race turned out to be strongly predictive.
Now, nearly two-and-a-half years in, it's the most reliable thing I've ever worked with.
It isn't perfect, and no model is, but it's been profitable month after month in a way no tipster I followed in my first decade of betting ever came close to matching.
Ryan and I still speak every Monday morning, and most of those calls are about what the system got right rather than what it got wrong.
By the end of 2024 I trusted the model enough to treat it as a business rather than a hobby.
From the first of January 2025, every selection, stake and return went into a formal record I could show to anyone who asked.
That first year finished with every month in profit, from £3,140.25 in January to a high of £7,150.75 in August.
The quieter months, like £4,050.60 in December, were still bigger than anything I'd managed in a good month before the model came along.
By the end of the year, I'd also stopped opening the email nervously and started treating it like the weather forecast, something you act on without arguing.
It was the first calendar year of my betting life in which I never had a losing month.
But it was 2026 where things really began to move.
At the turn of the year, Ryan made the biggest upgrade since the original build.
He added what he calls a travel feature, which measures how far each horse has journeyed from its yard to the course and how it's performed after similar trips in the past.
Some trainers send horses the length of the country and win, while others rarely get one to run its race after four hours in a horsebox.
The system had never been able to see that before, and it made an immediate difference.
January 2026 brought £5,310.40, comfortably the best January the model has ever produced.
March was bigger again at £6,700.55, as the turf Flat season got going and the model had more races to work with.
April was the best month of the spring at £7,240.80, with 37 winning bets from 58.
In May, Ryan added a second upgrade that tracks horses changing trip, because a step up from a mile to ten furlongs can bring out improvement the form book hasn't priced in yet.
May held its ground at £6,890.30, even though the going dried out fast and a lot of trainers kept their horses at home.
June came in at £5,420.60, the lightest month of the year so far, yet still more than I used to make in a whole season.
Taken together, the first half of 2026 came to £37,243.20, which is more than I'd made from betting in the whole decade before the model existed.
Ryan still updates the AI with every new race result, so the learning never stops.
Each month it gets a fraction more precise, and each quarter the interaction effects become a little more refined.
That's the advantage of a machine learning system over a fixed method, as it keeps evolving instead of standing still.
It doesn't back a horse because it won last time out and looks like it should win again.
It won't steer clear of a horse just because the price eased overnight.
It looks at the raw numbers and asks one question, which is whether the price on offer is meaningfully bigger than the horse's real chance of winning.
If the answer is yes by a wide enough margin, it flags the bet, and if not, it ignores the horse however obvious a winner it might look.
That kind of self-control is impossible to keep up by hand over months and years.
The AI never gets carried away after a winning streak, and it never chases losses after a bad week.
It runs the numbers and reports what it finds.
Over the summer of 2026, things went into overdrive.
July, August and September produced winning months of £6,430.55, £7,749.20 and £8,150.75, the last of which is the biggest month the model has ever had.
September's £8,150.75 came from a busy month of big-field handicaps, the kind of races where the conditions layer finds the most to work with.
I had more free cash than I'd ever had before, and my friends had noticed long before then.
My mate Dave was first in, back in June, having watched me turn a profit month after month before finally asking if he could copy my bets.
I said yes, on the condition that he kept a full record and gave me his real assessment after 30 days.
His wife sent me a bottle of whisky.
By the end of the summer, Dave had made a profit for four months in a row.
He wasn't following every selection, because he was cautious in the way most people are when watching someone else's system before going all in.
But even with his conservative approach, he was banking between £800 and £1,400 most months.
Before long, his brother-in-law asked Dave to forward the bets to him too.
The brother-in-law, a plasterer called Mick, rang me after his first fortnight to ask whether the email ever had a bad day, because he hadn't seen one yet.
I told him it does, and that the system is judged across a month, never on a single afternoon.
I let a few more people in on the same terms, which meant following the selections, keeping their own record and giving me feedback after 30 days.
All 6 of them finished in profit, with the lowest at £1,200 and the highest at over £3,000, and not one of them lost money.
I kept their results alongside mine, and their monthly totals tracked my own closely, which told me the selections didn't rely on anything only I could do.
That's when I began thinking seriously about opening this up more formally.
Once you scale to hundreds of members, the effect on the market becomes a real problem, because your selections move the odds before half your members have logged in.
I've watched it happen to other services, where strong early results bring in more members and the advantage wears away because too many people follow the same output.
It's a manageable figure, spread across enough accounts and bookmakers that the market doesn't feel it.
It's small enough that the odds hold up when the email goes out each morning, yet large enough that I know I'm helping people.
I spent a couple of weeks putting this page together, not to sell something I was unsure about, but because I'm certain about what I have and want to share it with the right people.
I've written it the way I'd explain it to a friend in the pub, without the hype I used to see on the pages that took my money.
People who've been around betting long enough to know the difference between a reliable system and a lucky streak.
People who are ready to act on what the data shows rather than second-guess it.
"How many tipsters can you follow before you stop believing any of them? For me it was seven, and Bet Compiler Pro is the eighth and the only one that's put £1,965 in my account inside a month."
Rodney Chalmers, Stockport
"Day 1, two selections on before half seven and a winner by teatime; day 30, £2,310 up and I haven't missed a morning. The only change to my routine is that I now check the results on my lunch break instead of the news."
Hazel Pritchard, Carlisle
"I used to give up every Saturday afternoon to the form book and count myself lucky to finish level. Now it's eight minutes on a weekday morning, and my first month finished £1,780 in front."
Stuart Ellwood, Durham
"Matthew, you told me not to judge it on the first week, and you were right, because week two alone paid my golf club fees for the year. I'm £2,460 up after 30 days and my brother has already asked me for the link."
Brenda Cottrell, Taunton
I keep a record of every bet and every monthly total.
Nobody told me to, but in that bookmaker's office the traders who were any good kept meticulous records.
You can't improve what you can't measure.
The figures for Bet Compiler Pro come straight from my own accounts.
Since January 2025, the model has produced 22 months of profit in a row, with October adding £7,592.50.
This year alone, January to October added up to £67,166.20.
That's £13,460.80 more than the same ten months of 2025, with the best still possibly to come.
That's 22 consecutive months in profit.
Over £101,000 has been withdrawn in that time.
They aren't professional gamblers, and few of them came from a betting background.
One member is a retired lock-keeper from Gloucestershire who'd never placed a bet online until his son-in-law walked him through it.
Another is a secondary school teacher who places her bets on the school bus on the way in.
A third is a semi-retired accountant from Ludlow who reads the email at the golf club and places the bets before he tees off.
If the model's output is clear enough for them to act on, it's clear enough for anyone.
It's about following a structured output from a machine that doesn't guess.
Each morning the AI analyses every race and builds a selection list based purely on where the mathematical value sits.
There's no opinion involved and no soft spot for famous trainers or fashionable yards.
Nor does it lean towards the horse that won well last time and looks certain to win again.
Those are the human biases that cost punters money year after year, and my model has none of them.
It processes the data, spots where the price and the probability part company, and flags the bet.
If there's nothing worth betting, it flags nothing.
Occasionally, when conditions don't offer clear value, the email will tell you to sit it out.
Most human tipsters can't hold that line, as the urge to send something out and give members their money's worth for the day is enormous.
I've seen it ruin a lot of good tipsters, who have a quiet week and start forcing selections to look busy.
My model has no ego and feels no need to justify itself by producing bets.
It produces bets when the evidence supports them, and that's why it's stayed in profit for 22 months running.
As I write this in late October, the season is turning in a way that catches a lot of punters out.
The ground is softening by the week and the jumps season is gathering pace.
Form from the summer suddenly counts for much less, and horses that looked unbeatable on fast ground in August are now turning out on heavy going at Chepstow and Wetherby.
This is the time of year when the conditions layer earns its keep, since it knows how each horse's profile shifts when the ground changes.
Several of those winners came in races where the favourite had only ever run on good ground, and the model saw that coming a mile off.
Days like these are when I'm most glad the system doesn't share my old habits, because the 2024 version of me would have kept backing summer form well into November.
October is also when Ryan runs his twice-yearly review of the feature weights, ready for the winter programme.
This year he's added extra weight to a horse's record after a long break, because so many jumpers return from a summer off.
"Retired electrician, 64, and not one for fancy words: I made £1,320 in my first month. Every bet took less time to place than my morning shave."
Barry Holdsworth, Wakefield
"My dad has never once let me help him with anything technical, but he let me set up his Bet Compiler Pro emails. He finished his first month £1,590 up, and now he's the one explaining interaction effects to me."
Joanne Ridley, Lincoln
"I managed a betting shop in the Midlands for eleven years, so I know what a winning customer looks like from the wrong side of the counter. I joined in August, I was £2,880 up by mid-September, and my old area manager will never hear about it from me."
Geoffrey Mold, Hereford
"It's like having a friend in the weighing room who only rings when there's something worth hearing. After a month I'm £1,705 better off, and not one bet took longer than the wait at my bus stop."
Rosemary Fenwick, Morpeth
Entry is controlled for reasons that have nothing to do with marketing.
When the same selections go out to too many accounts, the market reacts and early movers can shift the price.
If all my members piled into the same horse within 20 minutes of each other, the odds would shorten before everyone got on.
That would wear away the very advantage you're paying for, so 30 memberships keeps the effect on the market manageable.
As I write this section, 21 of the 30 places have already been taken.
That leaves 9, and while you're still reading this page, one of them is yours.
When the last place goes, I'll take this page down and start a waiting list, and I can't say when that list will next be opened.
On the morning of your first day as a member, you'll receive an email with the day's selections.
Each one will include the horse, the race, the course, the time, and the price you should be looking to get.
Placing the bets takes most members under 10 minutes.
The email is plain and short, with no adverts and no long essays to wade through.
If there are no selections that morning, you'll still get a message saying so, and you're never left wondering.
You can do it online with any major bookmaker or take it to a shop if you prefer, whichever suits your routine.
There's no need to watch the races unless you want to, and no late-night form study.
The system handles all of that between one selection email and the next, and you act on what it sends.
Members have told me they didn't realise they had winners running until they checked their account after dinner, which shows how hands-off this is.
The morning email carries the load, and all you do is place the bets.
I've got retirees in my group alongside students and busy professionals, and every one of them finds the time to follow along.
One member is a hospital consultant who places his bets between ward rounds.
Another is a retired joiner from Pershore who'd never bet online before joining.
If either of them can do it, so can you.
You don't need to understand how the model works any more than you need to understand how your car engine works to drive it.
The selections land in your inbox and you back them, and over the weeks the balance grows.
Most members take under 10 minutes each morning to go from opening the email to having their bets placed, and after that the day is yours.
If you can open an email and type a horse's name into a bookmaker's search box, you have every skill this needs.
These are the questions that have landed in my inbox most often since I started letting people in, along with my answers.
The model finishes its overnight run at around 6.30am, and the email is usually with members by 7.15am.
There's no rush, because most races run in the afternoon, so placing the bets over breakfast or at your morning break works fine.
On the rare morning when the going changes overnight, the model waits for the updated reports and the email goes out a little later.
Because I watched that settlement software do in one second what took me five minutes, and it never made a mistake.
It taught me that a well-built program beats a tired human at repetitive, number-heavy work.
Weighing up 60 inputs for every runner on a busy card is that kind of work, done 30 or 40 times before breakfast.
No, and that's my settler's head talking.
Every multiple stacks the bookmaker's margin on top of itself, which is why the shops love accumulators so much.
The selections are built to be backed as singles, and that's how each figure on this page was made.
Any bookmaker can restrict any winning account, and I won't pretend otherwise.
The best protection is to spread your bets over two or three accounts and use a shop now and then.
I share the habits that have kept my own accounts open for over two years with every new member.
It has no favourites, but it does send more selections from courses with bigger fields and deeper historical records.
Small tracks with tiny fields give it less to learn from, so it treats them with more caution.
That's entirely your choice, and it should always be money you'd be comfortable seeing lost on a bad day.
Most early members started small and only raised their stakes once a full month had gone by.
Most weeks, because eight years of settling left me with strong opinions about horses.
I've learned to keep them to myself, as the AI's record over 22 months is better than mine ever was by a mile.
My 9 years of records cover British racing, and that's what the model was trained on.
Irish racing has its own patterns and smaller markets, so I won't send selections there until Ryan can show me the model works on it.
Yes, and a couple of members do, mainly because exchanges don't tend to restrict winning customers.
Just remember the commission comes off your winnings, so compare the exchange price with the bookmaker price before you place each bet.
Ryan has a share of the arrangement we agreed on that three-hour phone call.
He's also proud of it, because the tennis framework he built at university has turned into the most successful thing he's ever made.
No, it's a one-off payment, and nothing else will ever be taken from your card for this membership.
The AI scores whatever racing there is, but fewer races mean fewer opportunities, so those emails are often shorter.
Some of the best-priced winners this year have come on quiet midweek cards that most punters ignore.
Each selection comes with a short line explaining the main reason it was flagged, such as a favourable trip change or a strong finishing effort last time out.
Members tell me that line is the most interesting part of the email, as they're learning to read races the way the model does.
The service you join today won't be the same service in six months' time, because it'll be better.
The selections you receive in spring 2027 will come from a model that has processed everything that happens between now and then.
That includes every winner and loser, every market movement and every going report.
It all feeds back in and the system adjusts, which is the compounding advantage of a system that learns.
The longer it runs, the more refined the output becomes.
You're not just buying into today's performance, you're buying into a system that improves with time.
That isn't £20 a month, it's £20 once, with no subscription and no renewal.
Here's the reasoning behind that price.
I could charge a monthly fee like most services, but a subscription would turn this into a business I'd have to feed with new members every month.
A one-off £20 keeps things small and manageable, and it means the people who join are here for the results rather than a trial they'll forget to cancel.
My accounts are performing well, but anyone who's run a winning system for a while knows that spreading the bets across more accounts keeps things running smoothly.
30 members placing these bets across 30 separate accounts, none of them linked to mine, spreads the activity in a way that protects the advantage.
It's a practical arrangement that works for both of us.
I run multiple betting accounts to keep my winnings away from their automated flag systems.
On weekends, I love to bet in my local town centre, choosing a new shop every time.
So far, it's worked, and my AI has generated up to £8,150.75 in a single month.
30 new members placing these bets across 30 different accounts keeps everything spread out and unremarkable to the bookmakers, which benefits both of us.
Your membership comes with a full 30-day money-back guarantee that allows you to follow every selection.
Watch the results come in and decide for yourself whether Bet Compiler Pro meets the standard you expect.
You have the entire 30-day period to use the service as intended and see the bets play out in real time.
You can judge the accuracy, track the results and feel the confidence that comes from a clear, structured approach.
If at any moment within those 30 days you feel that the service isn't for you, you can request a refund and you'll receive 100% of your joining fee back without delay.
There are no questions asked and no pressure.
You won't need to fill in a form or explain yourself, as a short email is all it takes.
Nobody has taken me up on that refund yet, but it's there in black and white and I'll honour it every time.
If you leave this page now, I can't promise it will still be here when you return.
The remaining places will go when they go, and once they're gone the service will be closed to new members until further notice.
If you're ready to take this step, click the button below and complete the checkout for just £20.
You'll have access to today's selections before you've finished reading your confirmation email.
P.S. The morning email is probably landing right now for today's members.
In a few minutes they'll have their bets on and their day will carry on as normal.
The 30-day guarantee means you're carrying no financial risk at all in that first month.
The £20 joining fee is less than a single modest stake on most of the selections the model sends.
Several members were in profit for the month before the end of their first week.
Even if your first few days are slow, the guarantee covers you for the full month.
Take the full 30 days to follow every selection and track every result.
If you're not satisfied, you get your £20 back and we part ways without any awkwardness.
Yours sincerely,
Matthew Wilkinson