Separate support from valuation
Knowing a roster well and estimating its chance accurately are different skills. Following a favourite team can help you notice a role change, but it can also make a memorable clutch feel more representative than an ordinary loss. Treat this as a possible error in your process, not a diagnosis of every fan. The practical question is whether your estimate changes when the names are visible. You can examine that question with a written forecast made before checking the price and with a consistent record of what information changed your mind.
A hypothetical pricing example
The following prices and probabilities are invented to explain asymmetric margin. They are not an audit of a named team or an operator. The assumed favourite probability is 0.7042. Calling a quote shaded only has meaning relative to that assumption. Observing two different prices in the market cannot reveal the true probability or prove that public money caused the difference. A margin-removal method is another model, so a disagreement between proportional and additive estimates is sensitivity to assumptions, not a measurement of a fan tax.
ASYMMETRIC MARGIN ALLOCATION: PUBLIC FAVORITE vs UNDERDOG Assumed probabilities for this example: P(Favorite) = 0.7042 -> fair price = 1 / 0.7042 = 1.420 P(Underdog) = 0.2958 -> fair price = 1 / 0.2958 = 3.381 Posted prices: Favorite: 1.30 implied = 1 / 1.30 = 0.7692 Underdog: 3.40 implied = 1 / 3.40 = 0.2941 Total overround = 0.7692 + 0.2941 = 1.0633 (6.33% book margin) Where the margin actually sits: Favorite EV per unit = 0.7042 x 1.30 - 1 = -0.0845 (-8.45%) Underdog EV per unit = 0.2958 x 3.40 - 1 = +0.0057 (+0.57%) Naive proportional de-vig of the posted prices: Fav no-vig = 0.7692 / 1.0633 = 0.7234 -> 1.382 Dog no-vig = 0.2941 / 1.0633 = 0.2766 -> 3.615 => Proportional de-vig UNDERSTATES the favorite tax (1.382 vs true 1.420) and makes the underdog look worse than it is (3.615 vs true 3.381). The whole 6.33% margin, and then some, is charged to the favorite.
If the offered price equals the fair price multiplied by (1 − s), where s is the shade fraction, then the expected value per unit staked is exactly −s. Proof: EV = p × fair × (1 − s) − 1, and since p × fair = 1, EV = (1 − s) − 1 = −s. Shading 1.42 down to 1.30 is an 8.45% price cut, so every unit bet at that price is expected to lose 8.45 cents. Your read on the team does not change this. Even if the brand-name team wins, you were paid at a price that loses money across the population of similar bets.
In this example, high win frequency does not rescue a short price. The assumed favourite wins about seven times in ten but needs 76.923077% at 1.30 to break even. The underdog at 3.40 has conditional EV +0.572%, using p = 0.2958. That small positive number disappears if the estimate is even slightly too optimistic. Therefore the example does not justify automatically opposing famous teams. The relevant comparison is probability multiplied by the executable price, with costs and uncertainty included, for either side.
What repeated pricing means
Suppose, only as a model, the same probability and odds apply to 200 independent bets. With a 2% fraction of the current bankroll, expected wealth is multiplied by (1 + 0.02 × EV)^200. Expected log wealth instead adds 200 times p ln(1 + 0.02(d − 1)) + (1 − p) ln(0.98). Exponentiating that log expectation gives a geometric-growth benchmark, not an exact finite-sample median. A realised drawdown also depends on the order of results. Do not label a projected terminal loss a measured maximum drawdown.
A blind review that preserves context
For a blind exercise, label the teams Alpha and Beta and assign the labels independently of their names. Build the input vector from data available before the match. The example below assumes six standardised differences and six weights; the resulting composite is shown explicitly. These are chosen teaching values. No particular three-month window or set of weights is claimed to be optimal.
Hypothetical inputs; no fitted historical dataset is claimed
D = .25*.70 + .20*.60 + .15*.35 + .15*.25 + .10*.30 + .15*(-.45)
D = .3475
k = 1.9
p_Alpha = 1 / (1 + exp(-k*D)) = .659316545247
p_Beta = .340683454753
EV_Alpha(1.30) = -.142888491179
EV_Beta(3.40) = +.158323746160These weights and the logistic scale are arbitrary teaching inputs. The table is not a validated model and the output is not a calibrated probability. In a real evaluation, overlapping measures such as first-kill conversion and 5 v4 conversion can double-count information. Fit transformations and coefficients on training data, then assess calibration on later held-out matches. Keep the actual roster, opponents, map version and available information time consistent. Remove names and logos only where doing so does not delete information your forecast legitimately needs. A blind score built on the wrong roster is still wrong.
Lock the initial forecast before revealing the names and odds. After the reveal, separate factual corrections from preference: a confirmed substitute can justify a new input; liking a star player cannot. Save both versions and the reason for the revision. Do not claim the process was blind if distinctive statistics already identified the team. A second reviewer can help with anonymisation, but is not a guarantee against shared modelling errors. The purpose is to expose changes in judgment, not to manufacture a reason to bet the opponent.
Exclusion and an honest audit
If matches involving a team repeatedly lead you to ignore limits or revise probabilities without evidence, exclude those matches in your written plan. The exclusion can cover both sides and related totals, since betting against your team does not remove the emotional connection. Choose a review date rather than an arbitrary promise that twelve months cures bias. A breach is a reason to pause and reassess the process, not to impose a recovery bet. Watching a match without a financial position remains an available way to support a team.
Audit forecast calibration, price achieved, realised profit, costs and rule violations separately. CLV against a documented closing reference is useful price evidence, but it does not prove either profitability or the removal of fandom bias. Group differences can arise from odds, competition, sample selection or uncertainty. Do not declare a psychological diagnosis from two averages. Keep the classification rule fixed before collecting the next sample, include skipped forecasts, and report how many observations each group contains. The actionable result may be a narrower betting scope or no betting on favourite-team fixtures.
A review should also allow an inconclusive answer. If the favourite-team group contains only a few matches, retain the uncertainty rather than inventing a success rate. Record a revision only when you can name its factual trigger. Repeatedly changing the classification or comparison window after losses can make any process look successful. Keep the earlier version so that the next review can distinguish improved forecasting from a more convenient explanation of the same results.
