PitchQuant is a public project for analysing football odds and historical match data. Its central engineering choice is to let deterministic Python scripts do the arithmetic while an LLM follows a prescribed sequence of rules and evidence checks. The project’s author says the workflow covers roughly 227,000 matches, but the public release does not include the raw dataset needed to reproduce its large-sample findings. Its backtests are therefore project-reported results, not independently verified evidence of future accuracy or betting profit.
What PitchQuant is designed to do
PitchQuant is an analysis workflow, not a model that simply asks an LLM to name a winner. The project describes a pipeline that starts with data acquisition and probability calculations, then applies movement and scenario rules, league or international submodels, direction and goal analysis, score estimates, consistency checks, and archiving. The repository labels its public release v1.2 and its core model V3.5.76; those are project version labels, not independent assessments of the software.
As an Amazon Associate I earn from qualifying purchases.
The project article frames its question as what can be learned from football odds and describes the market as noisy. Its phrase “Football is chaotic, markets are efficient” is the author’s framing, not an established guarantee about every league or market. Read the project article.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Repair Windows errors before they cause bigger problems3Fix the driver behind crashes, sound loss and screen glitchesWhy split calculations from language-model work?
Odds analysis mixes arithmetic with decisions about which evidence applies. PitchQuant’s architecture attempts to keep those jobs distinct: code calculates values from inputs, while the LLM follows a constrained checklist and applies versioned rules. That can make the workflow easier to inspect than asking a model to calculate, interpret, and justify a forecast in one unconstrained answer. It does not, by itself, prove that the rules or input data are correct.
#1 Best Overall
| Question | LLM used as an unconstrained oracle | LLM used as a constrained runtime |
|---|---|---|
| Who performs arithmetic? | The model may calculate and explain in the same response. | Deterministic scripts perform calculations; the model follows the prescribed workflow. |
| Will the same input reproduce the same numerical output? | Not assured by the architecture alone. | The project assigns numerical computation to scripts, which can make those outputs reproducible for the same inputs and code. |
| Can a reader trace decisions? | Reasoning may be difficult to audit consistently. | Versioned rules, a generated checklist, consistency checks, and archived analyses are intended to leave a more structured trail. |
| Can the reported historical findings be independently reproduced from the public release? | Not established by the choice of LLM architecture. | Not fully: the project says the raw large-sample dataset is omitted from the public repository. |
The repository describes 414 checks for its v1.2 / core V3.5.76 snapshot, while the project article describes 238 checks. Those are source- and snapshot-specific counts, not one settled count for every version. The PitchQuant repository documents the scripts, rule and skill files, lookup tables, checklist, and analysis archive.
What the analysis calculates
The repository describes several complementary calculations rather than one universal prediction score. De-vigging adjusts bookmaker odds to remove the implied overround when deriving market probabilities. Poisson score modelling estimates score probabilities from goal-rate assumptions, with a Dixon–Coles adjustment for low-score outcomes. Kelly criterion calculations are used as a relative ranking signal, not a guarantee of value or a recommended stake.
Rank #2
Other components examine odds movement, market calibration, scenario rules, and score lookup tables distilled from historical analysis. The scripts calculate deterministic quantities; the LLM’s role is to work through the relevant rules and checks. A numerical output can be repeatable while still depending on assumptions, data quality, market selection, and the way a rule is defined.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteWhat data and competition coverage does the project describe?
The title-matched project article describes a six-month analysis spanning 227,000 matches. The repository and methodology document also cite 227,495 matches for particular analyses. Those counts describe different project summaries and backtests; they should not be treated as a single exact count covering every calculation. The article and repository provide the project’s account.
Rank #3
The repository names the Premier League, La Liga, Bundesliga, Serie A, Ligue 1, Champions League, and Europa League, plus a separate Nations League submodel. It says other leagues are not calibrated, so its claims should not be generalized to all football competitions.
Named data sources include Football-Data.co.uk, ClubElo, Understat, API-Football, odds-api.io, The Odds API, and public Chinese Sports Lottery odds. These sources do not mean every input is available for every match or analysis. The repository does not include the raw dataset or API keys; some live or supplemental data sources require credentials. Its example setup uses Python 3.10 or later, repository scripts, an odds input file, and keys where applicable. The project does not establish that any particular provider is required for every workflow.
Rank #4
What the backtests report—and what the figures mean
The methodology document describes a 70/30 time-based training/test split and says rules are intended to pass a paired significance criterion of p<0.05, as well as sample-size and rollback gates. These are the project’s stated evaluation criteria; the public materials do not independently establish that every implementation detail or source record has been audited. The project’s methodology and backtest document explains its approach.
| Reported result | What it describes | How to read it |
|---|---|---|
| 55–58% directional accuracy | Selected project summaries from the repository and article, published in 2026. | A project-reported range for selected analyses; it is not a single universal rate across all matches or markets. |
| About 30% top-two score hit rate | Strong-signal matches in the project’s article/repository reporting, 2026. | A conditional subset result, not an all-match exact-score rate. |
| 34.5% versus a 48.7% favourite baseline | The project methodology’s reported 30,000-match time-split test of its online-learning layer. | The layer underperformed that baseline by 14.2 percentage points in the described test; the project says it demoted the component to logging only. |
Direction accuracy, top-two score hits, and market calibration answer different questions. A direction result is not an exact-score result; a strong-signal subset is not the full match population; and performance against one baseline does not establish that a system beats a market in general. The project itself presents its figures as methodology validation rather than future performance.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Why the public results are not fully reproducible
The methodology document says the raw 227,000-match dataset was left out of the public repository because of its size and source terms. It therefore says the large-sample analyses cannot be rerun directly from the public release alone. Readers can inspect the project’s code, rules, documented methodology, and reported results, but cannot use that release by itself to reconstruct the underlying large-sample runs.
This distinction matters: a documented time split and significance threshold describe how a project says it evaluated rules, while independent reproduction also requires access to the relevant data and enough detail to recreate the implementation and inputs. The available public materials support describing PitchQuant’s design and its author-reported findings; they do not support presenting those findings as independently confirmed.
What this project can teach beyond its reported scores
- Make arithmetic inspectable. Assign calculations to ordinary code where inputs, formulas, and outputs can be checked, rather than relying on a language model to improvise numerical work.
- Keep evidence and metrics specific. State whether a figure measures direction, score ranking, calibration, a subset, or the entire tested sample; do not merge unlike metrics into a broad claim of predictive power.
- Use baselines and time-aware evaluation. A model result has more meaning when compared with a defined baseline under a documented split. Even then, conclusions depend on the chosen data and implementation.
- Retire weak components. The project’s account of demoting its underperforming online-learning layer illustrates that a pipeline should preserve negative results and be willing to stop using a component that fails its own test.
PitchQuant’s repository notice states: “FOR ACADEMIC & EDUCATIONAL USE ONLY. Use for gambling/betting is strictly PROHIBITED. NOT betting advice.” It also states that long-term sports-lottery expected value is negative. The reported historical analyses should not be treated as a way to select bets or infer guaranteed returns.
The Tool Desk
Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
