How every number here is calculated.

CLV is computed in implied-probability space, not odds space. American odds convert to an implied probability first, because odds themselves don't average or difference linearly — treating them as if they do is the most common way betting analytics goes quietly wrong.

implied(a) = a > 0  ?  100 / (a + 100)
                    :  |a| / (|a| + 100)

CLV% = (implied(close) − implied(entry)) / implied(entry) × 100

Worked example, computed on this page by the same function that grades your bets — you bet +150 and it closes at +130:

implied(+150)  = 40.00%
implied(+130)  = 43.48%

CLV% = (43.4783 − 40.0000) / 40.0000 × 100
     = +8.70%

Positive means you got a better price than the market settled at. The benchmark is Pinnacle's close, and only Pinnacle's. When Pinnacle never priced an event, the bet is left ungraded with a recorded reason — we don't substitute another book's close, because a CLV figure whose benchmark changes per bet isn't a measurement, it's a mood.

Raw ROI can't distinguish a steady grinder from someone who hit one 12-leg parlay. The Sharpe ratio divides your average return by how much that return bounces around, so consistency counts.

rᵢ         = profit(betᵢ) / stake(betᵢ)      per-bet return
mean       = Σ rᵢ / n
variance   = Σ (rᵢ − mean)² / (n − 1)       sample variance
volatility = √variance

Sharpe     = mean / volatility

The quality bands shown in-app are:

Sharpe > 0.30   excellent
Sharpe > 0.15   good
Sharpe > 0.05   mediocre
Sharpe > 0      poor
Sharpe ≤ 0      negative

Only settled won/lost bets count — pushes are excluded rather than scored as zero returns, since a push isn't an outcome your edge produced. Below three settled bets the figure isn't computed at all.

Risk of ruin answers one question: given your edge and how much you stake, what is the probability your bankroll reaches zero? It is computed two ways.

Analytical (everyone, including the free tier)

p       = win rate, clamped to [0.01, 0.99]
q       = 1 − p
b       = decimal odds − 1               net payout per unit
cushion = 100 / max(0.5, stake%)         units of bankroll

if p·b ≤ q:   ruin = 1.0                 no edge → ruin is eventually certain
else:         ruin = min(1, (q / (p·b)) ^ cushion)

Simulated (Pro and above)

A Monte Carlo pass runs 1,000 bankrolls of 500 bets by default, staking a fixed percentage each time and counting how many fall below 1% of their starting balance. The reported figure is simply that bust count over the run count. Risk levels are banded at 2%, 10% and 25%, and the recommended maximum stake is the largest whole percentage that keeps analytical ruin under 5%, capped at 10%.

The Bankroll Digital Twin projects forward by replaying your betting profile thousands of times. It runs 1,000 independent paths by default over the horizon you choose, and reports the distribution rather than a single headline number.

per day, per path:   simulate bets from your profile
                     (win rate, odds, staking / Kelly fraction)

outputs:
  percentile bands   p10 · p25 · p50 · p75 · p90   per day
  median final bankroll, plus p10 and p90 finals
  probability of ruin across all paths
  average maximum drawdown
  average longest losing streak

The percentile fan is the point. A median outcome on its own hides the fact that the 10th-percentile path — one bettor in ten, with identical skill — can look like a disaster over the same window. Variance is not a rounding error at these sample sizes.

Odds are captured from the sportsbooks we track every 30 minutes, around the clock, since March 2026. That is the whole archive — we don't imply years of history we don't have. Closing lines are stored per event so that CLV is graded against a real captured price rather than reconstructed after the fact.

The same archive powers the public CLV archive and the MLB closing-line study. Both are published so the methodology above can be checked against real output rather than taken on trust.