Sharkbetting Methodology: Exchange-Baseline Value Betting

A value bet is a wager whose offered price is higher than the true probability of the outcome implies. The arithmetic is trivial when the true probability is known. The hard part is that no one knows the true probability. Every value-betting tool in production is, at its core, an answer to one question: which price do you trust as the proxy for fair?

Three families of answer have emerged across the past decade. The first is consensus-baseline: average the prices offered by a set of bookmakers and treat the average as fair. This was the dominant approach in the early 2010s, when exchange volume in non-UK markets was thin and the only continuously priced source was the bookmaker grid itself. Consensus is easy to compute and rarely produces extreme readings, but it inherits any shared pricing model the contributing books use upstream. Eight books running variants of the same supplier feed are not eight independent reads.

The second is the proprietary fair-price model: train a model on historical results and have it output a probability per outcome. The output is fully internal, which is its appeal and its weakness. A bettor cannot audit a black box, and the same opacity that hides genuine edge also hides correlated mistakes. A model that misprices home favorites in cold-weather games across an entire season produces alerts that all share the same error.

The third is exchange-baseline: take the matched price on a betting exchange (Betfair, Smarkets) or a prediction market (Polymarket) as the proxy for fair. Exchange prices reflect actual money positioned by participants who can both back and lay an outcome. They clear continuously rather than being set at the open and adjusted reactively. Exchange-baseline became viable as a standalone methodology around 2014, when Betfair liquidity in major European football and US basketball markets consistently cleared the threshold at which the matched price stopped flickering in response to small bets. Polymarket expanded the approach to US election and political markets in 2024, and its 2025 sports rollout extended exchange-baseline coverage to markets where regulated sports exchanges have never operated.

Sharkbetting was built on exchange-baseline because the math is short, the inputs are publicly visible, and the failure modes (Section 5) are well characterized. The remainder of this paper specifies how the rating is computed, how it is graded against closing prices, and where the methodology should not be used.

The core rating is computed per outcome. Given a bookmaker decimal price p_book and an exchange decimal price p_ex, the outcome rating is defined as:

outcome_rating = ((p_book - 1) / (p_ex - 1)) * 100

The exchange price p_ex is itself the result of an aggregation step across one or more exchange feeds. Sharkbetting supports Betfair and Polymarket. A user can opt to compare against either feed in isolation or against the union of both. When the union is selected, the aggregation rule picks the best available price per outcome:

p_ex = max(p_betfair, p_polymarket)   // when both feeds carry the outcome

Every exchange charges a commission on net winnings, which means the price a bettor effectively faces on the exchange is lower than the displayed price. Sharkbetting applies a commission adjustment before the rating is computed. Given a commission rate c expressed as a decimal (Betfair default 0.02, Polymarket default 0.02 on winning shares):

p_ex_adjusted = 1 + (p_ex - 1) * (1 - c)

The interpretation of the resulting rating is mechanical. A rating of 100 is fair. A rating of 105 means the bookmaker price pays 5 percent more than the commission-adjusted exchange price does, per unit stake. Ratings below 100 indicate the bookmaker price is short of fair; in a long sample, betting at ratings below 100 is mathematically losing absent a separate edge (promotion abuse, mistake odds, soft-line books that cap small). Ratings above 100 indicate positive expected value, with the size of the edge proportional to (rating minus 100).

Two practical notes. First, the rating is reported to one decimal place. Intermediate computation runs at full floating-point precision; rounding only happens at the surface. Second, the rating is computed per outcome, not per market. A single match has three ratings on a 1x2 market (home, draw, away), each independent. There is no aggregate match rating. This is intentional: averaging per-outcome ratings blurs which specific selection a bettor should take, which is the only actionable answer.

3.1 Worked illustrative example

Setup. Consider a hypothetical Premier League match between two top-half teams. A bookmaker offers the home win at decimal 2.10. Betfair Exchange shows the same outcome at 1.95 with roughly 80,000 GBP of matched liquidity on the back side. Polymarket does not list the market.

Step 1, commission. Apply Equation 3.3 with the Betfair default commission of 0.02:

p_ex_adjusted = 1 + (1.95 - 1) * (1 - 0.02) = 1.931

Step 2, rating. Apply Equation 3.1:

outcome_rating = ((2.10 - 1) / (1.931 - 1)) * 100 = 118.15

Read. A rating of 118.15 indicates a strong value signal. The bookmaker is paying roughly 18 percent more per unit stake than the commission-adjusted exchange price does. The same calculation without the commission step would yield 115.79, which understates the genuine edge by approximately 2.4 rating points. In a sample of 600 bets at this rating threshold, closing-line value (Section 4) is the empirical check that the signal is real and not an artifact of stale exchange prices.

Closing-line value (CLV) is the percentage difference between the price a bettor took at entry and the price the same market closed at. Given an entry price p_entry and a closing price p_close:

CLV = (p_entry - p_close) / p_close * 100

CLV matters because it is the strongest known empirical predictor of long-term betting profit. A bettor whose median CLV across a large sample is positive will, with sufficient volume, be profitable. A bettor whose median CLV is negative will not, regardless of which individual bets won or lost. The relationship is mechanical: a bookmaker that consistently allows bets at prices the market later corrects upward is, by construction, paying out more than the market eventually agreed was fair. CLV is the way that mechanical relationship gets measured on real bets.

CLV is the primary KPI for the Sharkbetting rating, not the rating itself. A rating of 110 that produces zero average CLV across 600 bets is a broken rating, regardless of how confident the math looks. A rating of 105 that produces consistent +2 percent CLV across the same sample is a working rating. The rating is the input; CLV is the output that proves the input was meaningful.

4.1 Sample-size requirements

CLV is noisy at small sample sizes. A bettor logging 30 bets at an average CLV of +1 percent has not demonstrated edge; the sampling distribution at that size easily contains zero. The rating evaluation protocol in Section 7 uses a fixed sample of 600 bets per rubric cycle precisely because that size is large enough for a 0.5 percent average CLV difference between two ratings to clear statistical significance at p < 0.05. Below 300 bets, the per-tool CLV figures we publish should be read as directional only.

4.2 Per-tool measurement protocol

For each tool tested in a rubric cycle, the CLV protocol is:

1. Capture the bookmaker price at the moment the tool surfaces the alert. This is the entry price.
2. Pull the closing exchange price within 60 seconds of kickoff for sports markets, or within 60 seconds of official market close for prediction markets.
3. Compute Equation 4.1 per bet.
4. Take the median across the 600-bet sample. Report median, not mean, because outlier closing prices on thin markets can swing the mean by several percent in either direction.

Closing prices are pulled from the same exchange feed used for getDelta (Equation 3.1), to avoid a methodology mismatch between the rating and the grade. CLV is logged hourly during a rubric cycle and frozen on the first day of the next cycle. Earlier cycles are never retroactively rewritten.

Exchange-baseline is not universally correct. Below is a list of regimes in which the methodology either degrades or fails outright. Each bullet is honestly listed, and the failure conditions are stated with enough specificity that a reader can identify them without our help.

• Thin liquidity markets. When matched exchange volume on an outcome falls below roughly 5,000 GBP equivalent, the displayed exchange price reflects one or two casual orders rather than aggregate sharp opinion. A small order at an extreme price can drag the displayed price several ticks away from where the next matched bet actually clears. Under thin liquidity, the exchange price is more noise than signal, and ratings computed against it are unreliable. Sharkbetting flags liquidity context next to ratings and recommends a per-bettor liquidity threshold (commonly 20,000 GBP) as a filter. The methodology is silent on what should replace the rating in this regime; a careful bettor either skips the market or falls back on bookmaker consensus, with the understanding that consensus has its own limitations.
• In-play lag. Exchange prices update in milliseconds during live matches. Bookmaker prices update on a delay (typically 2 to 10 seconds, longer on some operators). A rating computed during in-play moments where the two clocks have desynced will read as positive value even when the bookmaker price is about to be pulled. Sharkbetting does not currently publish in-play ratings as a standalone product, and the CLV protocol explicitly excludes in-play bets. Any reader using the methodology to manually grade in-play tools should treat the resulting ratings as suggestive rather than dispositive.
• Prop markets with near-zero exchange volume. Player props (points, rebounds, assists), corners, cards, and most novelty markets do not list on Betfair or Polymarket at meaningful volume. Equation 3.1 cannot be applied if there is no p_ex to compute against. For prop markets, exchange-baseline is not a methodology, it is a non-answer. Sharkbetting excludes these markets from its core rating product and does not pretend a number can be generated where the underlying input does not exist. Tools that publish prop ratings against an exchange baseline are using either an internal model labeled as exchange-baseline, or are comparing against the few props that do list (typically star-player point totals on top NBA matches), and a reader should ask which.
• Match-fixing and integrity events. When a market is moving on inside information (suspected or confirmed match-fixing, late team-news leaks before they become public), the exchange price moves first because exchange participants include the parties acting on the information. Bookmaker prices catch up minutes or hours later. A rating computed during this window will read as positive value because the bookmaker price is genuinely off the eventual fair price. The bet is also one a regulated bookmaker is likely to void or limit. Exchange-baseline cannot distinguish a genuine sharp opinion from an integrity-driven move, and the methodology offers no protection against following sharp money into a fixed match.
• Commission-adjusted truths. Equation 3.3 deducts a fixed commission rate from the exchange price. In practice, commission tiers vary per user (Betfair Premium Charge, volume-based discounts, Polymarket fee-free promotions) and can move the effective fair price by half a percentage point or more across different bettors looking at the same screen. The default rate Sharkbetting publishes is the median rate a serious bettor faces, not a personal rate. Two bettors at different commission tiers should read the same rating and come away with different verdicts, and the methodology does not resolve that.
• Stale exchange quotes. Exchange feeds occasionally drop or freeze for short windows. The aggregation rule in Equation 3.2 picks the highest displayed price across feeds. If one feed has frozen at a stale quote that is higher than the live quote on the other feed, the rating is computed against the stale price and reads as inflated value. Sharkbetting filters stale quotes by a freshness threshold (price must be matched within the last 30 seconds), but the threshold is heuristic, not provable.
• Bookmaker stake limits. A rating of 110 on a market the bookmaker will only allow a 50 EUR stake against is a different product from a rating of 110 on a market the same bookmaker takes 5,000 EUR on. The methodology does not encode stake limits, and a tool that surfaces a high rating on a market with a low effective limit is producing technically correct but practically narrow signal. Soft-book identification and stake-limit estimation are downstream concerns the rating is silent on.

We list these failure modes because the value of a published methodology is in knowing when not to use it. Any exchange-baseline tool that does not disclose a comparable list is either omitting it or has not characterized its own failure modes. Both are reasons to read the resulting ratings with skepticism.

The three families introduced in Section 2 are not interchangeable. The table below summarizes the trade-offs.

A working bettor commonly uses exchange-baseline for the bulk of pre-match liquid markets and falls back to consensus on lower-tier leagues. Proprietary models are appropriate where the first two are absent, but only with documented CLV evidence that the model is doing better than chance.

Every price snapshot underlying a published rating is logged with a UTC timestamp, the bookmaker key, the exchange feed, the matched volume at capture, and the raw decimal price on both sides. Snapshots are written hourly into a partitioned store and frozen on cycle boundaries.

Methodology changes are tracked in a public changelog, versioned semantically. Material changes (refresh interval, default commission, supported exchange feeds, sample size) receive a numbered version. Cosmetic edits do not. The changelog is the audit trail; if a rating moves between cycles and the changelog has not, the move is from data, not methodology.

Raw data is shared on request. Email methodology@sharkbetting.com with a short note about intended use. We will return the price-snapshot CSV and the scoring sheet for the requested cycle, typically within 48 hours.

The standing test sample for the current cycle is 2,400 price snapshots collected between February 1, 2026 and April 28, 2026, sampled hourly across NBA, NFL, English Premier League, La Liga, Bundesliga, and UEFA Champions League. The sample is sized so a 0.5 percent edge difference between any two ratings clears p < 0.05 within a single cycle.

Exchange-baseline value betting is a methodology with a short core formula, a small set of well-characterized failure modes, and a known empirical grade in CLV. It wins on liquid pre-match markets and loses on thin markets, props, and in-play. Consensus-baseline wins where exchange volume is absent. Proprietary models win where neither pricing source is available, conditional on the operator publishing the back-test.

Sharkbetting publishes the formula, the inputs, the failure modes, the test protocol, and a changelog of methodology versions because the value of a methodology to a third party is in being able to audit it without our cooperation. The next planned addition is a 60-day live CLV ledger, updated daily, showing the median CLV produced by the rating across rolling 30 and 60-day windows. The ledger is targeted for release in Q3 2026 alongside Methodology v5.0.

Comments, corrections, and replication results are welcome at the address in Section 7.