Relevance Verified: 20-03-2026
Last updated: 31-03-2026
My work is in predictive modelling — specifically the application of statistical methods to sporting outcomes and the question of where genuine predictive signal exists versus where noise is being mistaken for pattern. It's a field where the gap between what most bettors believe they're doing (analysis) and what they're actually doing (pattern-matching on small samples contaminated by randomness) is large and consequential. The vocabulary in this glossary will not make you a winning sports bettor — nothing short of sustained edge development and disciplined record-keeping will do that. What it will do is let you think more clearly about the mathematics of prediction, the difference between a streak and a signal, and the honest limitations of even the best statistical models when applied to genuinely uncertain events.
What foundational casino and betting terms anchor everything else in this glossary?
These are the core concepts. They apply across all games and markets, and they are the building blocks on which every predictive model in sports betting is constructed.
| Term | Category | What it means | Modelling context | Notes |
|---|---|---|---|---|
| Expected Value (EV) | Probability | The probability-weighted mean outcome of a wager — the long-run average return across an infinite number of identical bets | EV = Σ(probability × payout). A positive EV bet exists when your model assigns a higher win probability to an outcome than the bookmaker's implied probability — the gap between the two is your edge | Sustained positive EV across hundreds of bets is what distinguishes a skilled bettor from a lucky one — the sample size needed to distinguish the two is far larger than most players assume |
| Overround (Vig) | Market Structure | The excess of summed implied probabilities over 100% — the bookmaker's embedded margin; the structural drag on every bet that a model must overcome to generate positive expected value | A standard −110/−110 spread market at 4.76% overround means your model must be calibrated well enough to overcome nearly 5% of drag per bet — the break-even win rate is 52.4%, not 50% | Strip the vig by dividing each outcome's implied probability by the total implied probability sum — the residual is the true market consensus probability before the operator's margin is added |
| Model Calibration | Modelling | The degree to which a model's predicted probabilities match observed frequencies — a well-calibrated model's 60% predictions win approximately 60% of the time across a large sample | Calibration is the first test of any sports prediction model — before checking profitability, check whether your predicted win rates match your actual win rates at each probability bin | Most amateur models are poorly calibrated — they overestimate probability for strong favourites and underestimate for underdogs, a bias that bookmakers price for and profit from |
| Sample Size | Statistics | The number of independent observations in a dataset — the single most important variable in determining whether a pattern in data reflects genuine signal or sampling noise | A bettor who goes 7/10 on spread bets may have an edge — or may have run hot on 10 random events. These are statistically indistinguishable at n=10. Meaningful signal requires n=500+ | The most expensive cognitive bias in sports betting is concluding that a model works after 20–30 bets. This is why record-keeping across hundreds of standardised bets is the only honest evaluation method |
| Closing Line Value (CLV) | Performance Metric | The comparison of your bet's odds at placement versus the market's closing price — consistently betting at prices better than closing is the best statistical evidence of genuine model edge | CLV is model-independent — it validates edge without requiring a large win-rate sample. If your 200-bet history shows average positive CLV, your model is pricing the market ahead of its efficient state | CLV works because sharp sportsbooks (Pinnacle, Circa) accurately reflect true probabilities at close — a bet at better-than-closing-line odds was statistically a value bet at the time placed |
| House Edge / RTP | Game Math | The operator's structural advantage — in casino games, a fixed mathematical constant; in sports betting, variable and potentially negative for a skilled model builder who finds genuine edge | Casino games have fixed house edges — no model can change them. Sports betting markets can theoretically be beaten by a model that outperforms the market's implied probabilities — a fundamentally different mathematical structure | The practical reality: fewer than 2% of sports bettors beat the closing line consistently — the rest are paying the overround plus their own estimation errors |
| Bankroll | Risk Management | Dedicated gambling capital — from a modelling standpoint, the parameter that determines how many bets you can absorb before a losing run eliminates your ability to realise your edge | Even a model with genuine positive EV will experience losing runs. A 55% win rate on even-money bets has roughly a 1-in-20 chance of a 10+ bet losing streak. Bankroll sizing determines survival through variance | 1–3% of bankroll per bet is the professional standard; higher stakes per bet accelerates variance-induced ruin even when the model has genuine edge |
| Variance | Statistics | The expected squared deviation from the mean — in a betting context, the statistical spread of outcomes around expected value; the reason short-term results are poor indicators of model quality | A model with 55% true win rate on even-money bets has σ ≈ 0.497 per bet. Over 100 bets, σ_total ≈ 4.97 bets — meaning a 1σ losing run produces a result of 50 wins, indistinguishable from a no-edge model | Variance is the reason models must be evaluated over hundreds of bets, not tens — and why record-keeping, not gut feel, is the only honest performance assessment |
| Wagering Requirement | Bonuses | The turnover threshold before bonus winnings become withdrawable; iGaming Ontario caps at 30x; relevant to modellers who attempt to extract value from sportsbook promotions | Bonus arbitrage (using a positive-EV model to clear WR requirements) is theoretically possible but rare in practice — most sportsbook bonuses restrict markets and minimum odds in ways that eliminate or invert the expected edge | Always calculate the expected cost of the wagering requirement before accepting a sportsbook bonus: turnover × (house edge per qualifying bet) determines whether the bonus is positive or negative EV to clear |
| KYC | Compliance | Know Your Customer — mandatory identity verification at all iGO-licensed operators before any withdrawal; government-issued ID, proof of address, sometimes source-of-funds documentation | Professional-level bettors are sometimes flagged for enhanced source-of-funds checks when withdrawals significantly exceed deposits over time — this is a FINTRAC / PCMLTFA regulatory requirement, not discrimination | Complete at registration — an enhanced KYC review on a large withdrawal is avoidable with pre-emptive documentation |
That sample size point is where I want to spend a moment, because it is the most consequential statistical concept for anyone who bets on sports and believes they have an analytical edge. In a population of 10,000 purely random bettors all staking on 50/50 propositions, approximately 9 of them will go 10-for-10 by pure chance. Those 9 people all believe they have a system. They don't. The only way to distinguish genuine edge from favourable variance over a short sample is mathematically impossible — you cannot. Over 500+ bets with consistent methodology, the signal starts to emerge from the noise. Under 100 bets, you know almost nothing about your true win rate.
Author's tip from Adrian Beck, Statistical Modelling and Predictive Analytics Consultant: "The Poisson distribution for goal modelling is not a magic edge — it's a starting framework. The model's value lies entirely in how well you estimate λ (the expected goals parameter) for each team. If you use raw historical goal averages, you're using the same data the market has already priced. The edge, if it exists, comes from better estimates of λ than the market has — typically through recency weighting, strength-of-schedule adjustment, and incorporating expected goals (xG) rather than actual goals, which have higher variance. The distribution itself is just the translation engine; the forecast quality lives in your λ estimate."What predictive modelling and statistical inference terms do informed Canadian sports bettors need?
| Term | Category | Definition | Application | Notes |
|---|---|---|---|---|
| Poisson Distribution | Statistical Model | A probability distribution modelling the number of independent events occurring in a fixed time interval, given a known average rate (λ) — the foundation of goal-based prediction models for football and hockey | P(k goals) = e^−λ × λ^k / k!. At λ=1.4 goals/team/game: P(0)=24.7%, P(1)=34.5%, P(2)=24.2%, P(3)=11.3%. Apply independently to home and away teams to build a score probability matrix | Poisson assumptions (goals are independent events) are known to be imperfect — momentum effects after goals and the Dixon-Coles 0-0 score underestimation are documented limitations requiring model adjustments |
| Bayesian Inference | Statistical Method | A probability framework that updates prior beliefs about an unknown parameter based on observed data, producing a posterior probability — formally: Posterior ∝ Prior × Likelihood | Example: prior belief that a team wins 50% of home games; new data shows 4 wins in 5 recent home games; Bayesian update produces a posterior that shifts toward higher win probability, weighted by sample size | Bayesian inference naturally prevents overreacting to small samples — the prior acts as a regulariser that pulls extreme observations back toward the population mean, exactly correcting for hot-hand bias |
| Regression to the Mean | Statistical Principle | The statistical tendency for extreme observations to be followed by observations closer to the population average — because extreme results are partially caused by favourable randomness that is unlikely to persist | A team on a 6-match winning streak in the CFL is more likely to be performing near their true quality level with good luck than to have genuinely improved — the streak partially reflects variance, not elevated ability | Betting the continuation of a hot streak without adjusting for regression to mean systematically overpays for recent form — the market already prices recent results; your model must add information the market hasn't incorporated |
| Hot Hand Fallacy | Cognitive Bias | The mistaken belief that a person or team on a winning streak has an elevated probability of continuing to win — the cognitive error of finding pattern in statistically normal variance | In casino games with independent outcomes (slots, roulette, virtual sports), the hot hand fallacy is an absolute error — there is zero autocorrelation between independent RNG outcomes. In sports, limited genuine momentum effects exist but are much weaker than perceived | Research (Gilovich, Vallone, Tversky 1985) demonstrated that basketball shooting streaks are largely consistent with independent binomial trials — the original hot hand study is one of the most replicated findings in sports psychology |
| Expected Goals (xG) | Sports Analytics | A metric estimating the probability that a specific shot attempt results in a goal, based on historical data about shots taken from similar positions, angles and situations — the expected value of that shot | xG is a better λ estimate for Poisson models than raw goals — a team that wins 3-0 but generates only 0.8 xG was lucky; one that loses 0-1 but generates 2.4 xG was unlucky. Future performance reverts toward xG, not actual goals | xG-based models outperform raw-goal models in calibration over large samples — the metric is widely available for EPL, NHL and MLS matches and is now standard input for serious prediction work |
| Elo Rating System | Ranking Model | A method for calculating relative team or player strength, updating ratings after each match based on the result relative to expected outcome — developed for chess, extended to NFL, NHL, NBA and international football | Elo naturally implements Bayesian updating — a surprise result (upset) moves ratings more than an expected result; the K-factor controls how fast ratings respond to new information | FiveThirtyEight's NFL Elo model was shown to outperform betting markets on roughly 1–2% of games — a small but statistically real edge when applied systematically over hundreds of games |
| p-value | Statistical Test | The probability of observing a result at least as extreme as the actual result, assuming the null hypothesis (no edge) is true — a small p-value means the result is unlikely to have occurred by chance | A bettor going 60-for-100 on even-money bets has p≈0.028 — that is, a 2.8% chance this occurred randomly with no edge. Significant, but not conclusive at 100 bets. At n=500, a 60% win rate has p<0.0001 | Conventional significance threshold p<0.05 is inappropriately weak for betting model validation — given multiple testing across many strategies, require p<0.01 minimum before concluding genuine edge exists |
| Selection Bias | Research Error | The distortion that occurs when the sample used to build or validate a model is not representative of the population the model will be applied to — including survivor bias, hindsight bias and publication bias | Backtesting a betting model on historical data suffers from selection bias — the model is implicitly fitted to that data. Out-of-sample testing on genuinely unseen data is the only valid performance measure | Most published "winning" betting systems are selection bias artefacts — the systems that happened to work on the historical period they were built on, published while the equally numerous failed systems went unrecorded |
| Overfitting | Modelling Error | When a predictive model captures the noise in training data as if it were signal — producing excellent historical performance but poor out-of-sample prediction; the most common failure mode for amateur sports models | A model with 40 parameters trained on 3 seasons of NHL data will look impressive in-sample — and fail immediately out-of-sample because it has captured specific variance patterns that don't generalise | Regularisation (LASSO, Ridge), cross-validation and out-of-sample testing are the standard countermeasures; the simpler a model, the less likely it is to overfit — parsimony has genuine predictive value |
| Strength of Schedule | Sports Analytics | An adjustment applied to raw performance metrics to account for the quality of opponents faced — a team with a 6-2 record against the five weakest teams is not equivalent to a team with the same record against balanced opposition | SOS adjustment is essential for early-season NHL and CFL modelling when raw records are small-sample and schedule-dependent — it is one of the most reliable sources of market inefficiency exploitable by quantitative models | Elo systems implicitly adjust for SOS through the rating update mechanism — beating a high-rated opponent moves your rating more than beating a low-rated one |
Author's tip from Adrian Beck, Statistical Modelling and Predictive Analytics Consultant: "Bayesian inference is the formal cure for the hot hand fallacy. When a team wins 7 of their last 10 games after a 50/50 historical win rate, a naive analyst calls this a hot streak and bets the continuation. A Bayesian analyst recognises that the observed 70% is partly real signal and partly variance — and that the posterior estimate, pulling toward the prior, is typically 60–65%, not 70%. The practical question is how much you trust your prior (the long-run base rate) versus how much weight you give recent results. Small samples should move your estimate less than your intuition suggests. This principle is worth C$100 to anyone willing to apply it honestly."
Why does sample size matter more than any other variable in sports prediction — and how does regression to mean affect betting decisions?
These two concepts are deeply connected. Regression to the mean is what happens when a phenomenon that was inflated by sampling variance returns toward its true value as more observations accumulate. Hot streaks, slumps, remarkable individual performances — all are partially composed of variance that cannot persist. The degree to which they are real versus variance-driven is precisely the question statistical modelling addresses. And the answer almost always requires more data than you have.
The confidence interval chart above is the mathematical argument against premature conclusions about betting systems. At n=10 bets, the 95% confidence interval around a true win rate of 55% runs from about 24% to 86% — meaning if you're observing 8/10 wins, you cannot distinguish that from a 55% true win rate, a 70% true win rate, or random luck. The bands only narrow meaningfully above n=100, and only become actionable for model validation above n=300–500. This is why professional betting analysts require at minimum one full season of out-of-sample results before concluding any model has genuine edge.
The Canadian context for this matters. iGaming Ontario's sportsbook market — launched April 2022 under Bill C-218's framework — now includes over 50 licensed operators, many offering the same sports markets. If you're building a model for NHL games (Toronto Maple Leafs, Ottawa Senators) or CFL wagering, the market is competitively efficient across the licensed ecosystem. Finding genuine edge requires either superior information (earlier access to injury news, better xG modelling) or consistently better probability calibration than the sharp market's closing line. Both are achievable. Both require substantial work and an honest statistical accounting of model performance over hundreds of bets.
Deposit limits, session timers and self-exclusion are required by AGCO at all iGO-licensed operators — including sports betting platforms at OLG. These tools are worth using regardless of whether you approach sports betting analytically: even positive-EV models go through losing runs, and a session limit prevents a bad run from accelerating into chase behaviour that abandons the model entirely. Set limits before your first bet. You must be 19+ in Ontario, BC and most provinces (18+ in Alberta, Manitoba and Quebec). ConnexOntario is free, 24/7, at 1-866-531-2600. For sports markets, odds comparison and responsible gambling tools, visit the OLG home page, or access your account to set deposit limits before beginning any betting session.
