How PaddocksEdge Scores Every Horse: The Model Explained

Most algorithmic racing services tell you what they produce. Fewer explain how. This article walks through the scoring model behind PaddocksEdge — what data it considers, how those inputs combine into a single conviction percentage, and why the release threshold exists.

If you already know the track record and want to understand what is generating those selections, this is the right place to start.

The Problem With Manual Form Study

Researching a race card properly takes time. Recent form, going preferences, distance records, trainer and jockey statistics, class changes, days off the track — that is before you factor in field size, pace, or breeding patterns on soft ground.

Most recreational bettors do not have an hour per race. So they shortcut. They read a columnist's view, check the favourite's odds, and back something on feel. That is not analysis. It is pattern-matching dressed up as research.

The PaddocksEdge model does not shortcut. It runs the full assessment on every runner, every day.

What the Model Actually Considers

The scoring model draws on six broad categories of data. Each contributes multiple individual signals — not a single headline figure.

Form patterns cover recent runs, finishing positions, and the direction of each horse's form. A horse finishing second, second, third is read differently from one finishing third, second, second. Trajectory matters.

Conditions match each runner to the race it is about to run. Going, distance, class, and weather are all factored against the horse's historical performance in comparable circumstances. A strong record on good-to-firm ground produces a different signal on heavy going.

Connections look at trainer and jockey form across recent runners. Stable confidence signals — the kind that show up in how a yard's horses have been running in the weeks prior — carry weight in the model.

Breeding draws on sire and dam patterns across thousands of historical races. This matters most for younger horses with limited form data, where pedigree provides a statistical prior.

Race context considers field size, class level, pace dynamics, and the relative strengths of the other runners. A horse does not run in isolation. Its score reflects the specific race it is entering.

Time since last run maps the days since each horse last raced against performance windows associated with peak fitness. A horse returning from a long absence is scored differently from one running off a recent sharp run.

The model tracks 196,633 horses across 669 UK and Irish tracks, drawing on 18 months of historical data.

How the Conviction Percentage Works

Once the model processes all six categories, it combines them into a single figure: the conviction percentage.

This is not a star rating or a qualitative label. It is a calibrated probability score that reflects how strongly the data signals converge on a runner. A higher conviction percentage means more of the model's independent dimensions are pointing in the same direction.

The key word is convergence. A horse with strong form but poor going conditions and a jockey on a cold streak will not score as high as one where form, conditions, and connections all align. The model does not reward partial signals.

You read the conviction score in seconds. The research behind it takes considerably longer.

The Release Threshold

Not every runner that scores well becomes a published selection. The release threshold is the filter that sits between the model's output and what you see.

Only runners that clear a minimum conviction level are released. This deliberately suppresses marginal signals. The result is fewer selections, not more. On a typical race day, the model publishes around five selections — not a list of twenty.

That selectivity is structural. It is not a claim about quality. It is a design decision that prioritises conviction over volume.

Why Pre-Race Logging Matters

Every selection is published and timestamped before the race starts. No selections are added, edited, or removed after the result is known.

This matters more than it might seem. A service that publishes selections without timestamps can, in principle, curate its record retrospectively. It does not have to do so deliberately — even unconscious selection bias can distort a track record over time.

The PaddocksEdge mechanism removes that possibility. The timestamp is the evidence. Outcomes are graded automatically when each race ends, with no human editing step between the result and the record.

The record writes itself.

As of the time of writing, the track record shows 54 top-three finishes from 79 settled selections across 18 race days — a top-3 strike rate of 68.4%. Those figures update daily. Check the live track record for current numbers.

How This Differs From Form Guides and Ratings Services

Timeform and Racing Post are genuinely useful. Timeform's ratings represent decades of accumulated expertise, and Racing Post's form data is comprehensive. That is not a criticism. It is just what they are: data and commentary tools.

They give you the ingredients. They do not cook the meal.

PaddocksEdge does not replace those resources for someone who wants to do their own analysis. What it does is run a structured, consistent scoring process across every runner and surface only the selections where the model's confidence clears a defined threshold. No editorial opinion. No columnist's instinct. Just the data, scored the same way every day.

For a broader look at how the two approaches compare, the PaddocksEdge vs Racing Post comparison covers that in detail.

What the Model Does Not Do

The honest answer is that no model predicts race outcomes with certainty. Horse racing involves variables that no dataset fully captures: a horse that scoped badly overnight, a jockey who takes an unexpected tactical line, a field that bunches at a crucial moment.

The model does not claim to eliminate uncertainty. It claims to process available data more consistently and completely than a recreational bettor can do manually. That is a narrower claim. It is also the right one to make.

A 68.4% top-3 strike rate across 79 settled selections is a meaningful figure. It is also a figure from a specific window of time. Past performance does not guarantee future results. Check the live record, watch it over time, and form your own view.

Evaluating the Model for Yourself

The seven-day free trial gives you full access to every selection, including the conviction percentage and the complete data breakdown per runner. You can see the model's reasoning before the race runs.

That is the right way to evaluate an algorithmic service. Not by reading a description of the methodology, but by watching it work in real time against live races.

The free trial details explain exactly what is included and how the trial converts. If you want a broader assessment of the track record before committing, the 2026 review covers 120 days of settled data.

You are not being asked to trust a description of the model. You are being given the data to check it yourself.