- Joined
- Jul 11, 2008
- Messages
- 1,954
- Reaction score
- 185
- Points
- 63
The reason this matters is straightforward. Match reports contain managerial information in narrative form. Prose. A journalist describing Ange Postecoglou's second-half adjustment in a 1-0 loss at a mid-block defensive side. A tactical analyst writing about how a manager responded to going behind with twenty minutes left. A post-match press conference transcript where the manager accidentally describes his defensive shape in response to a journalist's question. All of that information exists publicly. None of it is structured. And unstructured information in a database is useless - which is why most of it just sits in articles nobody returns to.
AI gives you a practical way to extract structured fields from unstructured prose consistently enough to build something usable. The word consistently is doing a lot of work in that sentence. You won't get perfect extraction every time. You'll get good enough extraction often enough to build a database that informs your analysis rather than replacing it.
What to Track and Why
Before the prompts, the field structure. The same lesson from the referee database applies here - design the database before collecting data, not after. The instinct is to start pulling information and figure out the structure later. That instinct produces unusable notes after twelve managers.
The fields that actually matter for betting decisions are as follows.
Default formation and shape. Not just the starting shape but the defensive and attacking shape separately, because most modern managers run different structures in and out of possession. A 4-3-3 in possession that becomes a 4-5-1 without the ball is a different tactical picture than a flat 4-3-3 throughout, and the difference matters enormously for total goals and Asian Handicap analysis.
Pressing intensity profile. The most useful version of this isn't a single rating but a split by opponent quality tier - high pressing against weaker opposition, mid or low block against stronger sides. Some managers run the same pressing structure regardless of opponent. Others adapt significantly. Which category a manager falls into changes how you model their fixtures against different calibre opposition.
In-game adjustment patterns. What a manager does when losing at half-time. What they do when winning with twenty minutes left. Whether they make early substitutions to change the game or wait until the last ten minutes. Whether their default response to conceding first is a tactical shift or a personnel change or both. These patterns are more stable across time than most bettors assume and they're almost never incorporated into pre-match pricing.
Substitution timing profile. Average time of first substitution by match context - winning, drawing, losing. Whether they use the early substitution as a tactical tool or prefer stability in the starting eleven. Whether they consistently make all five substitutions or leave the bench largely unused in certain contexts. This connects directly to the substitution pattern article's in-play analysis.
Opposition-specific adjustments. Whether this manager changes their fundamental shape or press trigger based on the specific opponent, or whether they impose their own game regardless. The managers who adapt to opposition are harder to model from historical averages. The managers who don't adapt are easier to model but will occasionally be caught out by opponents who've scouted them thoroughly.
Aerial and set piece orientation. Whether the manager sets up specifically to exploit or defend set pieces, which feeds directly into the set piece specialist absence analysis. A manager who has built his attacking structure around delivery quality is more affected by that specialist's absence than one who treats set pieces as secondary.
Press conference communication patterns. Whether this manager is genuinely informative in pre-match briefings or gives deliberately vague answers. Whether his injury updates are reliable signals or consistent obfuscation. This connects to the press conference analysis workflow and helps you calibrate how much weight to give his public statements.
Where the Source Material Comes From
Three categories, in order of signal quality.
Tactical analysis sites and long-form match coverage. The Athletic, Spielverlagerung for European football, specific club-focused tactical writers on Substack. These sources contain the most useful raw material because they describe specific tactical mechanisms rather than general impressions. A detailed tactical write-up of one match can produce a better database entry than five standard match reports combined. The limitation is coverage - the top six and the European elite get thorough treatment. A Championship manager in their first season might have three genuinely useful tactical pieces in the whole year.
Press conference transcripts. Underrated as a data source for the database. Managers who talk openly about their approach - and some of them do, more than you'd expect - give you direct self-description of their tactical intentions, squad usage logic, and in-game thinking. The limitation is that elite managers have learned to say very little of tactical value. The Championship and League One level is more informative precisely because the media scrutiny is lower and the managers are less media-trained.
Standard match reports and post-match summaries. Lower signal density than the two sources above. Useful for patterns that appear repeatedly across many matches - substitution timing, defensive shape against specific opponents - but not reliable for nuanced adjustment analysis. One match report telling you a manager switched to a back three in the sixty-fifth minute is a single data point. Ten match reports showing the same switch in the same scoreline context is a pattern worth encoding.
The practical reality for most managers outside the elite tier is that your source material will be thin. A League One manager will have fewer usable sources than Pep Guardiola has profiles written about a single press conference. The database entry will reflect that - fewer confident fields, more uncertainty flags, a clearer note that specific patterns need confirmation from further observation. That's fine. A database entry that honestly reflects sparse coverage is more useful than one that fills fields with guesswork.
The Two-Prompt Architecture
Same principle as the referee database. Two types of prompt, used for different purposes, never collapsed into one.
The first type is the bulk extraction prompt. You give it a match report or tactical piece and ask it to extract specific fields from the text. The output goes into your database fields directly. This prompt is narrow, specific, and deliberately restricted from drawing conclusions.
The second type is the qualitative synthesis prompt. You give it multiple entries about the same manager and ask it to identify patterns across them. This prompt is wider, interpretive, and explicitly invited to reason about what the patterns mean. It runs after you've built at least three to five entries on a specific manager, not before.
Running them in the wrong order is the most common mistake. Asking the model to synthesise patterns from a single match report produces speculation dressed up as analysis. The bulk extraction prompt run five times across five sources produces something worth synthesising.
The Bulk Extraction Prompt
When you have a source document ready - pasted directly into the conversation or uploaded as a file - use a structure like this:
"I'm building a tactical database on football managers. The following text is [source type - match report / tactical analysis / press conference transcript] covering [manager name]'s team in [competition, date approximately]. I want you to extract only the information that is explicitly stated or clearly implied in this text. Do not infer anything that isn't supported by the source material. For each field below, give me either the extracted information or 'not mentioned in this source' if the text doesn't address it. Do not fill blank fields with general knowledge about this manager. Fields: default formation and defensive shape; pressing intensity and trigger points; any in-game adjustments described; substitution timing or specific substitutions mentioned; any set piece or aerial orientation; squad rotation or selection logic if mentioned. For each field you do fill, include the exact sentence or section from the source that supports your extraction."
The "include the exact sentence from the source" instruction is the most important line in the prompt. It forces traceability - you can see exactly what the model is basing each extraction on, which lets you verify it and flag cases where the extraction has stretched the source material. Without it, the model will generate plausible extractions that drift from what the text actually says.
The "do not fill blank fields with general knowledge" instruction addresses a specific failure mode. LLMs have read enormous amounts of football content and have general impressions of well-known managers. Without this instruction, they will quietly fill blanks with their training data impression of the manager rather than flagging the absence of information in your source. You end up with a database that looks complete but contains untracked AI inferences that you'll mistake for sourced findings.
The Qualitative Synthesis Prompt
Once you've run the bulk extraction prompt across at least five sources and have populated database fields for a specific manager, you run the synthesis prompt on the accumulated entries:
"The following are structured database entries extracted from multiple sources about [manager name]. Each entry was extracted from a specific source document. I want you to identify patterns across these entries - specifically, which fields show consistent information across multiple sources, which show variation that might indicate context-dependency, and which remain sparse or unclear. For pressing intensity, identify whether the pattern suggests a consistent approach or opponent-quality adaptation. For in-game adjustments, identify whether any response pattern appears repeatedly and describe it specifically. For substitution timing, calculate an approximate average if enough data points exist. Flag any fields where the pattern from your synthesis contradicts what a generic description of this manager would suggest. Identify your uncertainty - if a pattern is based on two sources, say so; if it's based on eight, say that instead."
The "flag contradictions with generic descriptions" instruction earns its place. It specifically asks the model to surface findings that might be counterintuitive - the manager with an aggressive reputation who actually sits deep in away fixtures, the conservative coach who makes his first substitution earlier than almost anyone else in the division. Those contradictions are where the database earns its value. Generic knowledge would have told you the wrong thing. Your specific sourced entries are telling you something different.
Handling Thin Data
For most managers below the Premier League and Bundesliga tier, source material is genuinely sparse. Three decent tactical pieces per season is optimistic for a Championship manager. For League One and below, you might have one genuinely useful analytical source per season.
The honest answer to thin data is partial database entries that explicitly acknowledge their incompleteness. Better to have a field marked as "insufficient data - two sources, inconsistent" than to fill it with a best guess.
The practical workaround for thin coverage is extended match report usage. Standard match reports have low signal density per report, but if you're processing twenty match reports for a specific manager over a season, the substitution timing data accumulates reliably even when the tactical nuance doesn't. Substitution time is mentioned in almost every match report. Formation at kick-off is mentioned in most. Specific in-game adjustments are described in fewer. Nuanced pressing analysis appears in almost none. Your database completeness will naturally reflect that hierarchy.
For managers you follow in real-time - watching their matches regularly - your own structured observation notes become a primary source. The opposition research article described a note-taking format for this. Notes you take during a match, entered into the extraction prompt afterward, produce better database entries than secondary journalism does, because you're specifically watching for the patterns you want to track rather than reading a journalist's general account of the game.
Connecting the Database to Fixture Analysis
A manager database with no connection to an actual betting decision is a research project. The connection comes through two places in the series workflow.
The first is the opposition research workflow. When you're building a match-specific tactical profile for an upcoming fixture, you start by pulling the relevant manager entries from your database. The opposition research prompts then use that structured information as a foundation rather than starting from scratch. The difference in output quality is significant - an opposition research prompt that receives structured manager tendency data produces match-specific analysis. The same prompt without that context produces general team summaries.
The second is the fixture screening tool. One of your screening criteria can be: does this fixture involve a manager whose database entry shows patterns that create specific market implications in this matchup context? A manager who consistently sits deep against top-half opposition, in a fixture where they're facing a team in the top six, playing a match-up where the total goals line is set above two and a half. The database makes that criterion possible. Without it, you're relying on general impressions and reputation, which is what the market is also using.
Maintaining the Database
Two maintenance tasks, one weekly and one seasonal.
The weekly task is adding new entries when a manager you track produces a match or press conference worth recording. This doesn't mean updating after every game - it means having a low threshold for adding entries when something genuinely new appears. A manager making an unusual tactical choice against a specific type of opponent. A press conference that reveals something about squad selection logic ahead of a cup tie. A substitution pattern that contradicts his established profile. These are worth capturing when they happen rather than in a batch review.
The seasonal task is running the qualitative synthesis prompt on each manager's accumulated entries from the previous season and updating the summary fields accordingly. Tactical tendencies drift over a manager's tenure - the pressing intensity that defined his first season fades as the squad ages, the squad rotation logic shifts as he learns which players to trust in specific contexts. A database entry from three seasons ago without a seasonal update is actively misleading.
The hardest part of maintenance is manager changes. When a manager leaves and a new appointment arrives, you start a new entry from scratch rather than inheriting the previous manager's profile. And for the caretaker period between appointments, you add a specific caretaker entry that explicitly notes the tactical simplification dynamic - because the match-script implications of a caretaker fixture are different enough from a settled manager that conflating the two would produce bad analysis.
Anyway. The database doesn't have to be comprehensive to be useful. Ten well-sourced manager entries covering the competitions you actually bet on regularly will do more for your opposition research workflow than a broad shallow database covering fifty managers you rarely encounter. Depth on the managers who matter to your betting is worth more than width across the whole game.
FAQ
Can I use the same prompts for assistant managers who've stepped up as caretaker?
Yes, with an important modification. For a caretaker who was previously an assistant at the same club, you can extract information about their likely approach from press conference comments during their time as assistant, tactical analysis pieces that mention the assistant's specific role in training organisation, and any previous head coaching stints elsewhere. The bulk extraction prompt works for all of these. The synthesis prompt needs an additional instruction: "Note that this individual has limited head coaching experience. Weight observations from their head coaching role more heavily than inferences from their assistant role, and flag where you're drawing on assistant-role information to fill gaps." The caretaker article's tactical simplification point applies here - their initial approach as head coach will likely be a simplified version of the outgoing manager's system, which itself is useful information.
How do I handle managers who change their system significantly mid-season?
Create a dated entry structure - entries with a specific time range rather than a single unified profile. A manager who ran a high press in the first half of the season and shifted to a mid-block after a run of results has two distinct tactical profiles that shouldn't be averaged together. When running the synthesis prompt, add the instruction: "Identify whether there is evidence of a systematic tactical shift at any point in the entry timeline. If there is, describe the before-and-after profiles separately rather than synthesising them into a single pattern." The database then contains both profiles with a rough transition date, and you use the more recent profile as the primary working assumption while noting that the earlier profile may still appear in specific match contexts.
Is this worth building for Championship and League One managers given the thin source coverage?
Honestly - yes, more than for Premier League managers. The market's information processing in the Championship and below is slower and less thorough. A structured database entry on a Championship manager built from twenty match reports and four press conference transcripts already contains more systematically organised information than the market's pricing models are incorporating. For Premier League managers, the marginal value of your database is smaller because the market has already processed far more information about them. The database is most powerful where market information processing is weakest, which is exactly the territory where coverage is thinnest. The effort required is lower and the edge relative to the market is larger. Build the database for the competitions you bet on most, regardless of whether those competitions produce rich tactical journalism.
