methodology

Research positioning

PitchCopilot is a sports quantitative research and publishing platform. Our goal is not to produce unquestionable answers, but to bring event data, external pricing benchmarks, calibration models, human review and public performance into the same traceable process. Readers can see how the model formed its opinions and can check afterward whether those opinions hold up over time.

The platform serves two types of readers: FPL players need more robust event background and risk judgment; sports creators need structured narratives that can be quoted, explained, and traced back. All content is based on data research and does not constitute investment advice or actual asset arrangements.

Market implicit pricing

Market-implied pricing is one of the starting points for our approach. It collates public market signals into comparable benchmarks so that the model is not just talking about itself in internal data. PitchCopilot compares market-implied pricing to model distributions, checking for explainable differences in wins, draws, goal ranges, pace and uncertainty.

When the model deviates from the benchmark, we will describe it as actuarial premium rate or Divergence. However, the gap itself is not a conclusion. It also requires data justification, historical calibration and manual review support. If the gap comes from data gaps or lineup uncertainty, the tone of the report will be lowered or observations will be included.

Asset weight configuration

Asset weight allocation is a research language used to express the model's relative confidence in different directions. The platform uses virtual points and virtual capital pools to put core configurations, secondary configurations and long-term results into the same standardized framework. These units do not represent actual money, do not promise returns, and are not operating instructions.

Weights are usually formed from three types of signals: event data, external pricing benchmarks and calibrated model distributions. If the three are consistent, the core configuration may be more concentrated; if the signals pull each other, the configuration will decline, become a second choice, or only remain for observation. This layering lets readers know about confidence differences and also allows for a fair check on public performance.

Settlement and hit definition

Every traceable configuration requires a hit definition beforehand. The settlement rules will explain the source of the results, the validity conditions, the treatment of postponement or cancellation of the event, and how the core and secondary configurations are counted separately. PitchCopilot should not rewrite definitions after the game, nor should it only retain favorable results.

Once the event is complete, results will be updated to /performance. Hit rate answers "whether it meets the ex-ante definition", Brier score answers "whether probability confidence is appropriate", and virtual point research observes weights and fluctuations. Only by reading these indicators together can we judge whether the model is truly reliable.

Brier Score and Calibration

The Brier score measures how close a chance prediction is to the actual outcome. If the model gives a high probability for an event but the event does not occur, the score will reflect overconfidence; if the model gives a 60% chance of an event over a long period of time, the calibration curve will be closer to the ideal line. PitchCopilot uses Brier scores and calibration curves to check whether the model is just hitting in the short term, or whether it truly expresses reasonable probabilities.

Calibration also affects version iterations. If a version performs poorly in a specific league or high-confidence range, the team will adjust the weight, release threshold or manual review reminder. Version changes should preserve context so that readers will not be confused when comparing different periods.

Data sources and quality control

Sources influence every subsequent judgment. PitchCopilot will collate event information, team background, open market signals, game results and manually supplemented data, and try to retain traceable context in the report. If data is delayed, fields are missing, or pre-match information has not been confirmed, the model output cannot be considered a complete conclusion. Human reviewers need to check whether the material is sufficient to support the core configuration, and if necessary, demote the content to second choice or observation items.

Quality control also includes language and compliance. The platform uses sports quantitative models, calibration models and data research as its main positioning to avoid readers misunderstanding the purpose of the content. Each public report should be able to link back to methods, FAQs, guidelines and performance pages to form a closed loop that is readable, verifiable and settleable. These internal links are not decoration, but allow readers to follow the complete process from a single article.

Model version and reflection iteration

When public performance reveals model deviations, the team will iteratively correct them through versions. Corrections may include feature weighting, calibration methods, release thresholds, league tiering, or manual review checklists. Version changes need to preserve the context and let readers know that the old and new samples cannot be read directly.

The point of reflective iteration is not to claim that the new version is necessarily better, but to allow the hypothesis to be tested. If the new version improves the Brier score but the fluctuation of virtual points expands, it means that it still needs to be observed; if the stratification of core configuration and secondary configuration becomes clearer, it may indicate that the release threshold is more stable.

How readers can verify

Readers do not need to understand every model detail to do basic verification. First check whether the report explains the background of the data, then check whether the configuration has clear hierarchies, and finally check the public performance for similar samples. If the content has no method links, no settlement rules, and no version context, it does not meet PitchCopilot’s requirements for credible research.

Release pipeline

PitchCopilot's reports are published using human review. The model first generates candidate studies, which are then checked by administrators for sources, compliance statements, strength of tone, settlement definitions, internal links and readability. It will be made public after approval; if there is insufficient data, the release may be delayed or changed to an observation item.

This pipeline forms a closed loop: data collection, model research, manual review, public release, post-game settlement, performance feedback, and version iteration. Readers can follow each step through /guides, /faq, /about and /performance. If there are any issues with the information or content, please report via /contact.

This translation is for convenience only; the Traditional Chinese version prevails.