- Joined
- Jul 11, 2008
- Messages
- 1,954
- Reaction score
- 185
- Points
- 63
The gap between stated staking rules and actual staking behaviour is one of the most consistent and least discussed problems in serious betting. Not recreational betting, where the whole point is largely intuitive and nobody pretends otherwise. Serious betting - the kind where you've thought about Kelly fractions, where you've run the Monte Carlo simulation, where you have an explicit framework for sizing bets based on estimated edge and bankroll percentage. Even there, probably especially there, actual stakes diverge from model-recommended stakes under emotional conditions in ways that are systematic, predictable, and almost never formally tracked.
The reason it's underdiscussed is that it requires admitting something uncomfortable. Most serious bettors believe their staking is more disciplined than it is. Not wildly undisciplined - nobody running a serious operation is doubling stakes after every loss in a deliberate Martingale. But quietly, conditionally undisciplined in specific circumstances that happen to correlate with emotional state. After a losing run, stakes creep down as confidence erodes. After a winning run, they sometimes creep up as confidence inflates. On fixtures you follow closely and have strong opinions about, the stake reflects the strength of your opinion rather than strictly the mathematical output of your staking model. On markets you're less familiar with but bet anyway, stakes are sometimes inflated because the unfamiliarity makes the edge feel larger than the model recommends.
None of this feels like a discipline failure from the inside. It feels like contextual adjustment. Sometimes it is. The monitor is what tells you whether it is or isn't.
What the Log Needs to Capture
The staking discipline monitor runs on a log that captures both what you staked and what your model recommended you stake. Without both columns, you're tracking behaviour without a baseline, which produces no useful comparison.
The fields required are as follows. Date and kick-off time. Fixture and competition. Market type. Your estimated edge at time of bet placement - not retrospective, the estimate you had before the match. The model-recommended stake based on that edge estimate and your current bankroll. Your actual stake. The ratio of actual to recommended - this is the key derived field, the one that converts two numbers into a directly comparable signal. Recent run context at time of bet placement - your last five results, expressed simply as the number of wins. Time since last winning bet in days. Whether you had an existing position in the same fixture or adjacent markets. Whether the bet was placed within six hours of kick-off or earlier. And a brief notes field for anything contextually unusual.
The edge estimate field deserves specific attention. Many bettors have staking models that take edge as an input - Kelly-based systems require it explicitly, and even flat-staking systems implicitly assume a threshold edge below which you don't bet. If you're not recording your edge estimate at the time of placement, you can't distinguish between intentional stake variation - staking more because you estimated a larger edge - and behavioural stake variation - staking more because of emotional state independent of edge. Those are different things with different implications, and conflating them makes the log useless for behavioural diagnosis.
If you don't currently record edge estimates, start recording them now, even roughly. A three-point scale - high confidence edge, moderate confidence edge, marginal - is better than nothing. A specific percentage estimate is better than a three-point scale. The precision of the log determines the precision of the findings.
The Monthly Review Prompt
The log accumulates for a month before you run the primary analysis. Weekly is too frequent - the sample in any given week is too small for the conditional patterns to appear clearly. Monthly gives you enough volume that genuine behavioural patterns separate from noise.
The primary monthly prompt:
"The following is my staking log for [month]. Each entry includes my estimated edge, my model-recommended stake, my actual stake, and the contextual fields described. I want you to identify conditions under which my actual stakes systematically diverge from my model recommendations. Specifically, examine the following conditions: recent run context at time of placement - do my actual stakes diverge from recommended stakes differently after losing runs versus winning runs? Time since last win - is there a relationship between days since last win and the direction or magnitude of stake divergence? Time before kick-off - do stakes placed within six hours of kick-off diverge differently from stakes placed earlier? Competition familiarity - I've marked bets in competitions I follow closely versus competitions I follow less closely - is there a pattern in stake divergence between these groups? Estimated edge level - do I over-stake relative to recommendations at high confidence edges and under-stake at marginal edges, or the reverse? For each condition, calculate the average actual-to-recommended stake ratio and compare it to my overall average ratio. Flag any finding based on fewer than fifteen bets in a condition category as directional only. Do not draw conclusions about whether the divergences are good or bad - just describe the patterns."
The "do not draw conclusions about whether the divergences are good or bad" instruction is doing important work. Some stake divergences are justifiable - your model's edge estimates are themselves uncertain, and sizing up slightly when your confidence is genuinely higher is defensible. The monitor's job isn't to tell you all divergence is wrong. It's to tell you where and under what conditions divergence occurs, so you can evaluate those specific cases rather than the general principle.
The Conditional Stake Ratio - Your Single Most Important Number
The actual-to-recommended ratio is what converts raw stake data into a diagnostic. A ratio of 1.0 means you staked exactly what your model recommended. Above 1.0 means you over-staked relative to the recommendation. Below 1.0 means you under-staked. The directional pattern of that ratio across different contextual conditions is what the monitor is measuring.
A ratio that is stable across conditions - roughly 1.0 in losing runs and winning runs, after long dry spells and short ones, in familiar competitions and less familiar ones - indicates genuine staking discipline. The model recommendation is doing the work it's supposed to do.
A ratio that varies systematically with conditions is behavioural evidence. Post-losing-run ratios below 0.8 indicate confidence-driven under-staking - you're betting less than your model recommends because recent losses have made you doubt your edge estimates even when nothing in the analysis justifies that doubt. Post-winning-run ratios above 1.2 indicate confidence inflation - you're betting more than the model recommends because recent wins have made the edge feel more certain than it actually is. Both are forms of stake variation driven by emotional state rather than analytical content.
The most interesting conditional pattern for most bettors is the relationship between time since last win and stake ratio. A log that shows stake ratios declining progressively as the dry spell extends - not dramatically, but consistently - is documenting the specific mechanism through which losing runs become extended losing runs. The under-staking during a dry spell doesn't cause the losses to continue, but it does mean you're betting smallest when your model's edge estimates have been most recently validated by closing line performance, and betting largest when your confidence is highest - which isn't always when the edge is actually largest.
The Edge Estimate Calibration Check
A secondary analysis the monthly prompt should include, run separately from the stake divergence analysis:
"Using the same log, I want to check whether my edge estimates are calibrated. For bets where I estimated high confidence edge, what was the average CLV? For moderate confidence edge? For marginal edge? If my edge estimates are calibrated, high confidence estimates should correspond to higher average CLV than moderate, which should be higher than marginal. Identify whether this ordering holds, whether any tier shows systematically higher or lower CLV than expected, and whether there are conditions - competition, market type, time of placement - where my confidence estimates and CLV performance are most and least aligned."
This analysis serves two purposes. It checks whether your edge estimates are actually predictive, which is a standalone useful diagnostic. And it connects back to the stake divergence analysis - if your high confidence estimates correspond to higher CLV but you're over-staking at high confidence relative to your model recommendation, the over-staking might be partially justified. If your high confidence estimates correspond to similar CLV as your moderate estimates, the over-staking definitely isn't.
The two analyses together produce a more honest picture than either alone. Stake divergence without edge calibration can be misread as discipline failure when it might be appropriate adjustment. Edge calibration without stake divergence analysis tells you whether your estimates are accurate but not whether your behaviour reflects them.
Addressing What You Find
The findings will fall into roughly three categories and each requires a different response.
Systematic post-loss under-staking is the most common finding and the one with the clearest fix. The fix is not trying harder to follow your rules - willpower-based solutions fail under the same emotional conditions that produced the under-staking in the first place. The fix is a structural rule: after any sequence of three or more consecutive losses, you are not permitted to reduce stakes below your model recommendation without first writing a one-sentence justification for why the edge estimate has changed. The justification requirement creates friction between the emotional impulse and the behaviour, and friction is usually enough to prevent the impulse from acting without engaging the analytical layer.
Post-win over-staking is less common among serious bettors than under-staking but exists. The structural fix is the same in reverse - after a winning sequence, require a written justification for any stake above model recommendation before placing.
Competition-familiarity divergence - staking differently in competitions you follow closely versus loosely - usually reflects appropriate variation if the edge estimate field shows higher estimated edges in familiar competitions. If the edge estimates are similar but the stakes are higher in familiar competitions, the divergence is confidence bias rather than analytical difference. The fix is applying your competition onboarding criteria more rigorously before betting in less familiar territory - if you don't have the confidence to stake at model-recommended levels, you don't have the confidence to bet there yet.
Time-before-kick-off divergence - late bets sized differently from early bets - connects directly to the betting history leakage analysis findings on temporal patterns. If your late bets both underperform on CLV and show stake divergence from model recommendations, both problems have the same root: bets placed under time pressure without the analytical foundation that justifies the stake level. The fix is a pre-kick-off window rule - any bet within four hours of kick-off requires a specific written justification for both the bet and the stake level before placement.
The Structural Rule Document
Once the monthly review has produced findings and you've designed responses to them, the responses need to live somewhere you'll actually see them before placing bets. Not in a folder of old analysis documents. In your pre-bet routine.
A structural rule document is a single page - literally one page, not a sprawling reference guide - containing your staking model parameters and the specific conditional rules your monitor has identified as necessary. It gets read before any bet is placed. Not skimmed. Read.
The LLM can help maintain this document. After each monthly review, run a prompt like this:
"Based on this month's staking discipline findings, update my structural rule document. The current version is: [paste current document]. The findings from this month's review are: [paste findings]. Identify any rule that needs strengthening based on persistent violations, any rule that can be relaxed because the condition it addresses has not appeared in three consecutive monthly reviews, and any new rule that this month's findings suggest adding. Keep the document to one page. Rules should be specific enough to be unambiguous at the moment of bet placement - avoid rules that require interpretation under pressure."
The "avoid rules that require interpretation under pressure" instruction is the most practically important line in that prompt. A rule that says "be more careful about stake sizing when on a losing run" requires interpretation. A rule that says "after three consecutive losses, write one sentence justifying any stake above model recommendation before placing" doesn't. Interpretation under pressure is where rules collapse. Specificity is what makes them hold.
The Quarterly Pattern Analysis
Monthly reviews identify current conditions. Quarterly analysis identifies whether patterns are persistent or improving - whether the structural rules you've implemented are actually changing your behaviour or whether the same conditions keep producing the same divergences month after month.
The quarterly prompt:
"The following are three monthly staking discipline reviews covering [date range]. I want you to identify: which divergence patterns have improved across the three months, which have persisted unchanged, and which have appeared for the first time in the most recent month. For any persistent pattern - one that appears in all three months with similar magnitude - identify whether the structural rule designed to address it appears in my rule document and if so why it may not be working. For persistent patterns without a corresponding rule, flag the gap explicitly."
The "why the rule may not be working" instruction is the hardest follow-up in the whole monitoring system. A rule that's present and being violated suggests either that the rule is too vague, that the condition triggers in a way you haven't anticipated, or that the emotional force driving the behaviour is stronger than the friction the rule creates. All three require different responses, and the quarterly analysis is what surfaces which case applies.
Anyway. The staking discipline monitor is probably the least immediately exciting thing in this series - there are no clever prompts, no novel data sources, no elegant database architecture. It is a log, a monthly analysis, and a rule document. What it does is convert the part of betting that most people treat as a character question - am I disciplined or not - into a data question: under what specific conditions does my behaviour diverge from my stated rules, and by how much. Character questions are hard to improve. Data questions have answers.
FAQ
My staking model doesn't produce specific stake recommendations - I use a rough percentage range rather than a precise figure. Can I still run this monitor?
Yes, but with slightly lower precision. Use the midpoint of your recommended range as the reference figure for the actual-to-recommended ratio. A bet at the top of your range produces a ratio above 1.0, a bet at the bottom produces a ratio below 1.0. The pattern analysis still works - systematic above-range staking in specific conditions and below-range staking in others is behavioural evidence regardless of whether your model outputs a precise number or a range. The finding that you consistently bet at the top of your range after winning sequences and the bottom after losing sequences is meaningful whether the range is 1.5-2.5% or a precise Kelly output. The precision matters less than the consistency of the conditional comparison.
What if the monthly review finds no significant stake divergence - does that mean my staking is disciplined or that the log isn't capturing the right conditions?
Both are possible and worth distinguishing. If your actual-to-recommended ratios cluster tightly around 1.0 across all conditions and your CLV record is positive, the most reasonable conclusion is genuine staking discipline. If your ratios cluster tightly around 1.0 but your CLV record shows markets or conditions where your process is underperforming, the log may not be capturing the relevant conditions. The conditions listed in the monthly prompt - recent run, time since win, competition familiarity, time before kick-off - are the most common sources of divergence but not exhaustive. Add conditions specific to your own betting patterns: matches involving teams you support, fixtures with significant media coverage creating public narrative pressure, markets where you have an unusually strong prior opinion. The log captures what you tell it to capture. If the generic conditions produce clean findings, the specific ones are worth testing.
How do I handle the edge estimate field honestly when my edge estimates are themselves uncertain?
Record the estimate you actually had at time of placement, including the uncertainty. "High confidence, approximately 4% edge" is a legitimate entry. "Uncertain - somewhere between 2% and 6% depending on team news that hadn't resolved" is a legitimate entry. "Marginal, probably below my threshold but placed anyway" is the most valuable entry you can make - it's an honest record of a bet placed outside your stated criteria, which is exactly the kind of finding the monitor exists to surface. The temptation is to record edge estimates that justify the stakes you actually placed, which converts the log from a discipline tool into a rationalisation archive. The monitor only works if the edge estimate reflects what you actually thought before you placed the bet, not what you needed to have thought to make the stake look correct.
