Why Do Positive EV Bets Lose? The Variance Math

Your +EV bets lose because expected value is a property of a distribution, not of a bet. A 3% edge at a near-even-money price moves your win probability from 50% to about 51.5%. You still lose 48.5% of the time. Forever. No model fixes that, because there is nothing to fix — the losses are the edge, working exactly as specified.

That is the whole answer to the question. Everything below is the arithmetic for the follow-up question people actually mean when they ask it: how long do I have to sit inside that distribution before my bankroll curve starts to resemble my spreadsheet, and how do I tell a real edge from a lucky one before I have 10,000 bets logged?

The number that governs your P&L, and it isn't edge

Everyone quotes edge. Almost nobody quotes variance, and variance is the term that dominates for the first several thousand bets.

Take a flat unit at decimal 2.00 with a true win probability of 51.5%. EV per bet is +0.03 units. Standard deviation per bet is very close to 1.0 units, because every outcome is either +1 or −1 and the mean sits near zero. So your signal-to-noise ratio per bet — edge divided by per-bet standard deviation — is 0.03.

That is a brutally small number. After n bets, cumulative profit has mean 0.03n and standard deviation 1.0 × sqrt(n). Signal grows linearly, noise grows with the square root. Set them equal and you get the point where your expected profit merely equals one standard deviation of noise:

0.03n = 1.0 × sqrt(n)
sqrt(n) = 1 / 0.03 = 33.3
n ≈ 1,111 bets

At roughly eleven hundred bets, a one-sigma unlucky run lands you at exactly break-even. Which means something like one bettor in six with a genuine 3% edge is still down at that point. Not down a hair. In the red, having done everything right, staring at a log that says the model doesn't work.

Away from 2.00, per-bet standard deviation is sqrt(p(1-p)) × odds, which for the 1.70–2.50 range most +EV plays live in lands somewhere between 0.9 and 1.6 units. Use 1.0 as a working default and your conclusions won't change.

Long run is a count, not a calendar

The phrase "long run" gets used as if it means "eventually, in time". It doesn't. It means a count of independent outcomes, and three things routinely wreck the naive mapping from bets placed to samples collected.

Correlated bets are not separate samples. Forty player props on the same NBA game, all keyed off one pace projection, is not forty samples. It's closer to three. If the game turns into a rock fight, they die together. Your effective n is far below your bet count, and your drawdowns will be uglier than the flat-bet model predicts. This is the single most common reason someone with a "1,500-bet sample" is really working from a few hundred.

Edge drifts. Markets sharpen. Data sources degrade. The specific soft book you were beating tightens its lines after a promotion ends. You are estimating a moving target with a slow instrument, and the instrument needs more samples than the target sits still for.

Staking rewrites the distribution. Kelly turns additive variance into multiplicative variance. Full Kelly on a 3% edge gives you something like a one-in-three chance of halving your bankroll at some point before it doubles. Quarter Kelly costs you roughly 44% of the growth rate and cuts drawdown severity by far more than that. Take the quarter Kelly. Bankroll survival dominates growth rate whenever your edge estimate has error bars — and yours does, which is the entire point of this post.

When does your measured ROI mean anything

The useful question isn't "when do I go green". It's "when does my observed ROI carry information". Standard error of observed ROI after n bets is approximately sigma / sqrt(n), with sigma ≈ 1.0 for flat units:

Read the first row again before you fire your model. Most people who announce that their +EV strategy "stopped working" are looking at 300 bets. That's a coin flip with extra steps.

At 20 bets a day — aggressive but achievable for someone scanning prematch markets seriously — 10,000 bets is about 500 days. So: how long until EV betting is profitable? Plan on eighteen months to know, not eighteen weeks.

The metric that converges faster than money

Closing line value. It is the only real answer to the sample-size problem, and it's underused mainly because it requires a reference price most bettors don't have on hand.

Here's the mechanism, and the mechanism is the whole reason it works. Your P&L contains two independent sources of randomness: did you get a good price, and did the ball go in. CLV discards the second one entirely. If you consistently take 2.10 on outcomes whose no-vig closing price is 2.02, your edge is real and it is measurable in a few hundred bets instead of a few thousand, because you deleted result variance from the measurement.

The reference has to be sharp. Grading yourself against a recreational book's own closing number tells you nothing except that the book moved. A de-vigged Pinnacle close is the standard reference because that book takes size, moves on sharp money and prices thin — its closing number is the market's best guess, not a marketing position.

Every market row across our REST and stream surface carries nvp, the no-vig fair price, so you aren't doing two-way normalisation by hand and getting it subtly wrong. A minimal tracker:

from pinnodds import Client

c = Client(api_key="YOUR_KEY")

def log_clv(bet):
    # bet = {"event_id": 1583, "market": "total", "designation": "over", "taken": 2.10}
    detail = c.details(event_id=bet["event_id"])
    close  = find_line(detail, bet["market"], bet["designation"])
    fair   = close["nvp"]                  # no-vig closing fair price
    return (bet["taken"] / fair) - 1.0     # > 0 means you beat the close

Worked example. You take Over 2.75 in a soccer match at 2.10. At kickoff the same line's nvp reads 2.02. Your CLV on that bet is 2.10 / 2.02 - 1 = +3.96%. The match finishes 1–1 and you lose the bet. Nothing about that result is evidence against you — you bought a 2.02 asset for 2.10 and the coupon didn't pay. Do that three hundred times with a mean CLV near +4% and tight dispersion and you have an edge, whatever the P&L column says.

The inverse is the uncomfortable one: mean CLV of 0.0% alongside a green P&L is a hot streak with a story attached. It will end.

One detail that matters more than it looks: capture the closing line at the right line. If you bet Over 2.75 and grade against the 3.0 close, you've measured the wrong instrument. Every price Pinnacle publishes comes through the feed, quarter lines included — 1.75 / 2.25 / 2.75 / 3.25 on totals, −0.75 / −1.25 / −1.75 / −2.25 on handicaps — so you can always match your bet to its own line rather than the nearest whole number. Depth is identical on every plan; tiers differ on rate limit and push access (pricing).

The honest caveats

The maths above is not the whole problem, and pretending otherwise would be a disservice.

Positive EV against the wrong benchmark isn't positive EV. If you compute your edge against a soft book's own de-vigged line, you've measured that book's margin, not your advantage. The edge has to come from a price a sharp market disagrees with. Get this wrong and no sample size rescues you — you'll converge, cleanly and patiently, on a small loss.

Stake limits cap the entire exercise. A 4% edge on $50 is $2. The bottleneck for most +EV bettors is not the model, it's how much they're allowed to put down before the account is restricted. The limit field in our data exists precisely because max stake carries information: a line accepting a large limit has been stress-tested by money, while one with a tiny limit is closer to an opening guess. Weight them differently in any model that consumes both.

A single reference book is a single point of failure. pinnodds carries Pinnacle. That's deliberate — it's the sharpest widely available reference, and it's the reason CLV against it means anything. But if your strategy is built on finding outliers across forty books, you need an aggregator alongside us, not instead of us. We're the benchmark, not the shopping list.

There is no historical odds archive. The feed is live and prematch, pushed in real time. If you want to backtest five years of closing lines, you have to start recording them today. Running a small consumer against /ws/feed that writes every tick to Postgres works fine and plenty of people do exactly that — but it's your job, not a product feature.

Drop alerts are a signal, not a strategy. The SSE stream at /odds-drop tells you a price fell against its own recent history. Often that's information. Sometimes it's a limit adjustment or a stale-line correction. Chasing every alert without a thesis is an efficient way to accumulate a large sample of negative-EV bets.

Instrument it properly, from bet one

Three fields, logged on every wager, no exceptions:

1. The price you took, with a timestamp.
2. The sharp no-vig price at the moment you took it (nvp).
3. The sharp no-vig price at close.

The gap between (1) and (2) is your claimed edge. The gap between (1) and (3) is your CLV. When those two disagree systematically, your model isn't wrong — it's late. You're finding moves the market has already made.

Live betting follows the same logic on a shorter clock. Prices move fast enough that polling misses the window outright, which is why the feed pushes rather than making you ask for it. If you're wiring up a drop consumer, note that the raw SSE alert and the REST row are different shapes on purpose: the alert names the market in sect, the moving selection in outcome, prices in from_price / to_price, the event in id, and carries limit and nvp — but no drop percentage. The /api/drops row is the enriched form, with market, designation, from, to, event_id and a precomputed drop_pct. Field-by-field detail is in the docs.

Takeaway

Stop grading yourself on profit for the first thousand bets. At that sample size the standard error on your ROI is several times larger than any edge you plausibly have, and treating it as feedback will make you abandon good models and reinforce lucky ones — the exact opposite of learning. Grade on CLV against a de-vigged sharp close instead. Log nvp at bet time and at close on every single wager, match each bet to its own line rather than the nearest round number, and only open the P&L tab once your bet count has four digits. If mean CLV is positive and stable, the money is an accounting formality. You just have to survive the variance long enough to collect it, which is the strongest argument there is for quarter Kelly and against the stake sizing that kills most people who actually had an edge.