- Joined
- Jul 11, 2008
- Messages
- 1,956
- Reaction score
- 185
- Points
- 63
The problem isn't that they're lying to themselves. It's that the picture they've assembled is almost always incomplete - and incomplete in a specific, predictable way. The summary statistics they track are the ones that confirm the broad shape of their activity. The leakage - the consistent negative patterns that sit underneath the headline numbers - stays invisible because nobody's looking for it in the right way.
That's what this article is about. Not building a CLV tracker from scratch (that's the previous piece). Not identifying cognitive distortions in your journal writing (that's the one after the CBT article). This is specifically about feeding your existing bet history to an LLM with prompts designed to surface structural leakage - the patterns in your own data that are costing you money, that you would almost certainly not find by staring at the spreadsheet yourself.
Why You Can't Find Your Own Leakage
Before getting into the prompts, it's worth being honest about why this analysis is hard to do without a structured approach.
The first problem is selective attention. When you review your own betting history, you're not processing it neutrally. You're processing it through the same cognitive architecture that placed the bets. Markets you prefer feel like they should be where your edge is. Competitions you follow closely feel like they should be where your analysis is sharpest. So when you scan the data, you naturally interpret ambiguous patterns in ways that confirm the story you already have about your own betting.
The second problem is the comparison baseline. A negative ROI in a specific market only means something if you know whether that market's win rate is consistent with break-even performance or genuinely below it. Most bettors don't have a clear baseline for each market type - they're comparing raw returns, which mixes together genuine edge variation, luck, and stake size inconsistency into a single number that's hard to interpret.
The third problem is that leakage rarely announces itself cleanly. It's usually conditional. Your Tuesday betting isn't worse than your Saturday betting in general - it's specifically worse in lower leagues on Tuesday evenings when you're tired and filling the card rather than analysing properly. That conditional pattern requires looking at the interaction between multiple variables simultaneously, and a spreadsheet isn't well suited to that without deliberate construction.
An LLM given a structured dataset and a well-designed set of prompts doesn't have any of those problems. It doesn't prefer your favourite markets. It doesn't have a story to protect. And it can identify conditional patterns across multiple dimensions simultaneously - if you ask it the right questions.
What Your Data Needs to Look Like
The quality of what you get back is entirely dependent on the quality of what you put in. A five-field spreadsheet and a thirty-field spreadsheet will produce completely different analytical outputs.
The minimum useful dataset for this type of analysis needs the following. Date and kick-off time - not just the date, the actual time, because day-of-week and time-of-day are both meaningful and the time is often the more important variable. Sport and competition - separate fields, not combined, because you want to be able to filter by each independently. Market type - match result, Asian Handicap, total goals, BTTS, player props, and so on as specific categories rather than grouped broadly. Bookmaker or exchange. Stake. Odds taken. Closing odds. Result. And a notes field, even if it's often empty - the presence or absence of a note is itself information about how much analytical work went into a bet.
The closing odds field is the difference between a useful dataset and a genuinely powerful one. Without closing line value, you can only analyse your actual results, which are heavily variance-contaminated over any sample you're likely to have. With CLV, you can analyse your process quality independently of luck - and that's where the most honest leakage detection happens.
If you don't have closing odds recorded historically, start recording them now. For the existing history, if you can approximate them from Oddsportal or similar historical data for your most common competitions, it's worth the effort. For the analysis below, I'll describe what's possible with and without CLV data.
The Baseline Audit Prompt
The first prompt isn't looking for leakage specifically. It's establishing the overall shape of your betting activity in a way that makes subsequent prompts more precise.
Give the LLM your full dataset as a CSV or pasted table, then use something like this:
"I'm going to give you my complete betting history. Before analysing it for patterns or problems, I want you to produce a factual summary of its structure. Tell me: the total number of bets, the date range, the breakdown by sport and competition, the breakdown by market type, the breakdown by bookmaker, the distribution of odds ranges I've bet into, and the distribution of stake sizes. Do not draw any conclusions yet. Do not identify patterns yet. Just describe the data accurately."
The reason for this sequencing is that you want the LLM to establish the structure before it starts analysing, because you'll use the structure summary to check whether subsequent findings are meaningful or are artefacts of small samples. A finding that your Championship betting has a -12% ROI looks different if it's based on two hundred bets versus if it's based on fourteen. You need the baseline to calibrate.
Once you have the baseline, note any surprises. Most people have a different mental picture of how their activity breaks down than the data shows. The competition distribution is often surprising. The market type split is often surprising. The bookmaker concentration is almost always surprising. These aren't findings yet - they're context for the prompts that follow.
The Leakage Identification Prompts
Now you're asking for the actual analysis. These should be run as separate prompts rather than combined, for the same reason the personal referee database prompts were separated - each question deserves the model's full attention, and combining them produces shallower treatment of each.
Market Type Performance With CLV Adjustment
"Using the betting history I've provided, analyse my performance by market type. For each market type with more than twenty bets, calculate my average CLV, my average ROI, and the relationship between the two. I want you to identify specifically: market types where my CLV is positive but my ROI is significantly below what positive CLV should produce over this sample, market types where my CLV is negative regardless of ROI, and market types where the CLV-to-ROI relationship looks unusual in either direction. Where a market type has fewer than twenty bets, note it separately and flag that the sample is too small for conclusions. Do not speculate about causes yet - just identify the patterns."
If you don't have CLV data: replace the CLV references with a request to separate results by odds range, since CLV-approximation by odds range gives you a rougher version of the same insight. High-odds markets with negative ROI tell a different story than low-odds markets with negative ROI, and separating them prevents the whole picture from being contaminated by a single market type that's performing badly.
Temporal Pattern Analysis
"Analyse my betting history for temporal patterns. Specifically, I want you to examine performance broken down by day of the week, by time of day (morning, afternoon, evening, late evening), and by whether the bet was placed more than 24 hours before kick-off, 6-24 hours before, or under 6 hours before kick-off. For each dimension where a meaningful sample exists, calculate average CLV and ROI. Flag any dimension where performance deviates significantly from my overall average, and note the sample size for each category. I'm not looking for explanations - just the numbers and which ones stand out."
The late-decision betting window - under six hours to kick-off - is where most people's leakage concentrates. Not always, but often enough that it's worth examining specifically. The late bet is usually either an injury-news reaction, a price movement chase, or a card-filling impulse. Only the first of those has analytical justification, and even then the edge depends on whether you're genuinely acting on information the market hasn't fully processed or chasing a line that's already moved past the value.
Staking Consistency Analysis
"Analyse my staking behaviour against any staking rules implied by the data. First, identify whether my stake sizes appear to follow a consistent pattern or vary significantly. Then, examine whether stake size correlates with odds - am I staking more on shorter-priced bets or longer-priced bets in ways that seem systematic? Next, look at whether my staking shows evidence of chasing - higher average stakes following losing sequences than following winning sequences. Finally, identify any competitions or market types where my stakes are systematically higher or lower than my overall average without obvious justification from odds level alone. Flag uncertainty where samples are small."
Most bettors believe their staking is more consistent than it is. The chasing signal is the most uncomfortable finding - not because it's surprising in the abstract but because seeing it in your own data, specifically, is different from knowing it's a general human tendency. The competition and market-level stake variation is often surprising too. Most people don't consciously stake heavier on their favourite league. They do anyway.
The Competition-Level Diagnostic
"Break down my performance by competition. For each competition with more than fifteen bets, calculate CLV, ROI, average stake, and average odds. Then identify: competitions where I'm betting with negative CLV consistently, competitions where my average stake is significantly above or below my overall average without being explained by odds level, competitions where I have high bet frequency but below-average CLV, and competitions where the pattern suggests I'm betting on familiarity or interest rather than genuine edge. Be direct about what the data shows. If a competition looks like it's costing me money systematically, say so clearly."
The familiarity-not-edge pattern is the most common competition-level leakage. You follow a league closely, you feel like you know it well, and that feeling of knowledge is hard to distinguish from actual analytical edge. The data usually knows the difference. A competition you've been betting for two years with a consistently negative CLV is telling you something your confidence in that league is refusing to hear.
The Cause Analysis - A Separate Step
Once you've run the identification prompts and have a clear picture of which patterns exist, you run a second round asking for possible explanations. The separation matters. If you ask for causes at the same time as patterns, the LLM will generate plausible-sounding explanations for everything it finds, including the noise. Separating identification from explanation forces it to commit to a pattern first and theorise second.
For each significant finding - say, consistently negative CLV in a specific market type, or deteriorating performance in the late-decision window - use something like this:
"I've identified that my CLV in total goals markets is consistently worse than in match result markets across the same competitions. Generate three or four possible structural explanations for this pattern. For each explanation, tell me what additional data or analysis would confirm or rule it out. Do not assume any single explanation is correct - treat them as hypotheses to test."
The hypothesis-testing framing is important. Without it, the model will land on the most plausible explanation and stop, which is exactly what you don't want. You want competing explanations evaluated against the evidence, not a single narrative that happens to fit.
What the Data Will Actually Tell You
This is worth being direct about before you run any of this.
The findings will probably be uncomfortable. Not catastrophically uncomfortable - more like the low-grade discomfort of discovering that something you believed about yourself is only partially true. The competition where you feel most expert is frequently not the competition where your CLV is strongest. The market type that feels most natural is frequently not the one producing the best process quality. The stakes you think you're controlling are frequently more erratic than you remember.
That discomfort is the point. The analysis that confirms everything you already believed about yourself is the analysis that isn't working.
There's also a sample size caveat that deserves more emphasis than most people give it. Serious leakage detection needs enough bets per category to produce signals rather than noise. A rough rule: treat anything under thirty bets in a category as directional indication at best, and reserve conclusions for categories with fifty bets or more. If your data is thin in important categories, the most honest output of this exercise might be a list of hypotheses to track going forward rather than confirmed findings to act on immediately.
Anyway. The goal isn't a comprehensive audit that produces a clean report. The goal is finding the one or two things that are quietly costing you money in ways you couldn't see from the headline numbers. One concrete leakage finding worth acting on is worth more than a thorough analysis of patterns that turn out to be noise.
Acting on What You Find
A finding without a specific change is just an interesting observation. Before finishing the analysis session, turn each significant finding into a concrete decision.
Negative CLV in a specific market type means one of three things: stop betting that market entirely, reduce stakes significantly until the CLV picture changes, or identify specifically what's driving the negative CLV and address it structurally. Vague intentions to "be more careful" in a market don't change anything. A specific rule - no total goals betting in League Two fixtures, for instance - changes the behaviour.
Temporal leakage findings translate into scheduling rules. If late-decision betting is costing you, the rule is that any bet placed within six hours of kick-off requires a specific written reason for why the late timing is justified. Not a vague sense that it still looks good. An actual written reason. The friction of writing it is part of the point.
Staking inconsistency findings translate into a staking audit habit. After every betting session, check whether your actual stakes matched your pre-stated sizing rules. Not weekly or monthly - after every session. The gap between real-time self-assessment and delayed review is where the inconsistency gets rationalised away.
FAQ
How much data do I need before this analysis is useful?
Roughly three hundred bets minimum to get meaningful results across most categories simultaneously. Fewer than that and too many individual category samples will be under the noise threshold to produce reliable findings. With fewer than a hundred bets, the most useful output is probably identifying which categories you need to track more consistently going forward rather than drawing conclusions from current data. If you're early in your tracking history, the value of this exercise is partly in clarifying what to record, not just in analysing what you have.
Should I run this analysis on a friend's data as well as my own?
There's a reasonable argument for it if you have a trusted betting contact willing to share. Looking at someone else's leakage patterns before examining your own produces a useful calibration - you're likely to notice conditional patterns in their data that their proximity to it prevents them from seeing, and the experience of doing that sharpens your eye for the same patterns in your own data. It also removes the defensive instinct. It's easier to identify a problem in someone else's betting history than in your own, and having done it once, the pattern recognition carries over. The reverse - sharing your own data with someone who will run the same analysis - is more uncomfortable and more useful still.
The LLM keeps finding patterns in small samples and presenting them confidently. How do I manage this?
Add an explicit instruction to every prompt: "For any finding based on fewer than thirty observations in a category, flag it explicitly as a low-confidence observation and do not include it in your main findings. Present it separately at the end as a pattern worth monitoring." Without that instruction, the model will present a -23% ROI in seven bets with the same confidence as a -8% ROI in two hundred bets, because it's pattern-matching rather than applying statistical intuition. The instruction doesn't eliminate the problem entirely but it forces the model to make the sample size visible, which lets you apply your own judgement about what to act on.
