People mix these up constantly. Arbitrage is betting both sides of the same event at prices that guarantee a profit no matter who wins. +EV is betting one side because your estimated probability is better than the price. One removes risk. The other uses risk. They are not the same strategy, and they should not share the same bankroll rules.

What Arb Actually Is

You see Team A at +110 at one book and Team B at +110 at another. Both sides cannot be fairly priced at +110 at once. If you stake the right amounts on each, you lock a small, certain return. No model required. The edge is mechanical: two books disagreed enough that the combined implied probabilities dropped under 100%.

That is different from middling (betting both sides at different numbers and hoping the result lands in the gap). Arb does not need a specific score. It just needs two prices that cannot both be fair.

What +EV Actually Is

You think a side is 54% to win and the no-vig market is pricing it at 51%. You bet that side only. You will lose plenty of individual bets. Over a large sample, the extra 3% is the profit. There is no locked outcome. Your edge lives in the quality of the estimate.

Why Treating Them as the Same Wrecks You

• Arb is tiny and fragile. Limits get cut fast, one side often gets voided or delayed, and the juice you "locked" can vanish if a book cancels. It also teaches nothing about whether your model is good.

• +EV needs variance. If you hedge every +EV bet into an arb, you are converting a larger expected profit into a smaller certain one, and you are paying extra vig to do it.

• Sizing is opposite. Arb stakes are set by the two prices so both sides pay. +EV stakes are set by bankroll and edge (Kelly or a fraction of it). Mixing those formulas is how people overbet noise.

A Simple Split

1. If both sides are already on the board at prices that sum under 100%, that is arb. Size to lock, if you even bother.

2. If only one side is wrong versus your number, that is +EV. Take the side. Do not "make it safe" unless the math of a hedge is clearly better than the remaining edge.

3. Grade them separately. Arb ROI is not evidence your model works.

Guaranteed money is not the same as expected money. Know which one you are chasing before you click.

Do you ever convert a +EV side into an arb, or do you let the variance ride?

A line is not "the market." It is that book's market. Sharp books try to post a number that is hard to beat. Square books try to post a number that recreational customers will bet into. Treating them as the same source is how people misread both value and CLV.

What "Sharp" vs "Square" Actually Means

A sharp book takes action from professional bettors, keeps limits higher, and moves quickly when informed money hits. Their number is often the closest public proxy for consensus true odds.

A square book makes most of its money from recreational volume. It will shade toward the public side (favorites, overs, star player props), hold a wider vig, and lag when the real market moves. The posted line can stay wrong longer because the customer base is not attacking it.

Neither is "better" in the abstract. They are built for different customers.

How This Changes Your Process

• Shop, but know who you are shopping. Getting +3 at a square shop vs. +2.5 at a sharp shop is real value. Using the square number as your benchmark for whether you beat the close is not. Track CLV against a sharp close, not against a soft book that never moved.

• Opens at square books are often the weakest number of the week. That is where a model can get paid. By close, many of those books have copied the sharp number and the edge is gone.

• Limits tell you the truth. A great price you can only bet $25 on is not the same opportunity as a slightly worse price you can bet $2,000 on. Edge times stake is what matters.

A Simple Rule

1. Use sharp books to learn what the market thinks.

2. Use square books to take a better price when they lag or shade.

3. Never grade your model against the book that was wrong the longest.

The job is not to beat every screen. It is to beat the honest number, at the book still offering the dishonest one.

Where do you take your closing-line benchmark from, a sharp book or an average of everything you can see?

Win-loss record is what the public prices. Pythagorean expectation is what the results should have been, based on points scored and allowed. Teams that win more games than their scoring margin supports tend to cool off. Teams that lose more than their margin supports tend to bounce back. That gap is one of the oldest, most durable signals in sports analytics.

The Core Idea

Bill James originally used it in baseball. The same logic shows up in football, basketball, and soccer: actual wins are noisier than the underlying scoring.

A common form:

Expected win% ≈ PF^x / (PF^x + PA^x)

PF is points for, PA is points against, and x is a sport-specific exponent (around 2 in baseball, often near 2.37 in the NFL, different again in basketball). You are not copying baseball. You are asking: given this scoring profile, how many games should this team have won?

Why Record Misleads Bettors

• Close-game luck. A team that is 8-1 in one-possession games is usually not an 8-1 talent. Those games are mostly coin flips. Pythagorean record strips a lot of that luck out.

• Blowout distortion, in both directions. A few huge wins can pad margin without meaning the team is that much better in a 3-point spread game. That is why some models use a capped or "Pythagorean with blowouts compressed" version.

• The market often knows this. Simply fading "lucky" records is not free money. Edge appears when the line still treats a lucky record as true talent, or when your exponent/sport fit is better than a naive win-loss read.

How to Use It

1. Compare actual record to expected record. Large gaps are a flag, not an automatic bet.

2. Combine it with opponent-adjusted ratings, not as a standalone system.

3. Recalculate on a rolling window so you are not mixing last year's team with this week's roster.

Record tells you what happened. Pythagorean tells you how sustainable it was. Bet the second number when the first one is doing too much work in the line.

Do you still start from wins and losses, or from scoring margin?

Most betting models start the same way: assign each team a power rating, convert the gap into a predicted margin, then add home-field and situational adjustments. If that rating is wrong, every number downstream is wrong. Elo, SRS, and similar systems are just different ways of answering one question: how good is this team right now, relative to the rest of the league?

What a Power Rating Actually Is

A power rating is a single number that summarizes a team's strength. The difference between two ratings is your expected margin (before extras like rest, weather, or travel).

A simple version:

Predicted margin ≈ (Team A rating − Team B rating) + home adjustment

If Team A is +4.5 and Team B is −1.0, and home is worth +1.5, you get a 7-point spread. That number is only useful if both ratings are measuring the same thing, on the same scale, with the same recency.

Where People Break It

• Win-loss ratings. A 10-2 team that beat a weak schedule is not a 10-2 team that beat quality. Ratings built on results without opponent adjustment systematically overrate padded records.

• Too slow or too fast. If you barely update after each game, your rating is last month's team. If you fully overwrite after one result, you are chasing noise. Elo-style systems solve this with a K-factor: how much one game is allowed to move the number.

• Mixing units. Points of rating, win probability, and spread are not interchangeable until you convert them. A 4-point rating gap is not a 4-point spread in every sport.

How to Use It Without Fooling Yourself

1. Rate teams on opponent-adjusted performance, not raw record.

2. Decide explicitly how much recent games weigh vs. full-season data.

3. Compare your rating-implied number to the market. The gap is the bet, not the rating itself.

A power rating is not a pick. It is the baseline. Edge lives in the difference between that baseline and the line, after you have already accounted for the stuff the rating does not see.

Do you rebuild ratings from performance data, or mostly from market prices?

A model that is always "confident" is not a model. Calibration is whether your stated probabilities match reality: when you say a side is 60% to win, it should win about 60% of the time across a large sample. Most bettors never check this. They only track wins and losses.

What Calibration Actually Measures

Win rate tells you whether you won. Calibration tells you whether your confidence was honest.

• A well-calibrated 55% bucket should land near 55%, not 70% and not 40%.

• Overconfidence looks like this: your "locks" win 58% of the time while you priced them at 70%.

• Underconfidence looks like the opposite: you keep calling something 52% when it actually hits 60%.

Either error is expensive. Overconfidence makes you size too big. Underconfidence makes you pass on real edge.

Why This Beats Raw Record as a Diagnostic

A 20-bet heater can hide a badly calibrated process. Calibration, plotted in buckets (50-55%, 55-60%, 60-65%, and so on), shows whether your probabilities are systematically off. You need volume in each bucket, not just an overall win rate.

If your 52% plays and your 65% plays both win at ~54%, you do not have a probability model. You have a coin with a slight bias, and you are labeling it with fake precision.

How to Use It

1. Log every bet with a pre-bet probability, not just a pick.

2. Group those probabilities into buckets and compare predicted vs. actual hit rate.

3. Recalibrate when a bucket is consistently high or low. Do not just "feel" more or less confident next time.

CLV tells you if you beat the market's number. Calibration tells you if your numbers mean what you think they mean. You want both.

Do you assign an actual probability to every bet, or just a side and a unit size?

Most bettors look at -110/-110 and think "that's basically a coin flip." They're looking at a number that already includes the book's tax. Devigging is the process of removing that tax so you can see the market's actual implied probability, not the price it wants you to pay.

What You're Looking At Before You Devig

At -110/-110, each side implies 52.38%. Those two numbers add up to 104.76%, not 100%. The extra 4.76% is the hold. If you treat 52.38% as "the true chance," you are comparing your model to a number that is already inflated in the book's favor.

The Simple Method (Proportional)

The most common approach is to scale both sides so they sum to 100%:

Fair probability = implied / (implied_home + implied_away)

For -110/-110, that gives 50/50. For a more lopsided line like -250 / +210:

• -250 implies 71.43%

• +210 implies 32.26%

• Sum = 103.69%

• Fair favorite ≈ 68.89%

• Fair underdog ≈ 31.11%

That gap between 71.43% and 68.89% is the tax you were accidentally treating as true odds.

Why This Matters for a Model

If your number says a team is 70% to win, and you compare that to the raw -250 implied of 71.43%, you think you have no bet. Compare it to the no-vig 68.89% and you actually have a small edge. The reverse happens too: a "value" bet against a raw implied can disappear once you strip the hold.

This is even more important on props and live markets, where the hold is often 6-10%+ instead of ~4.8%. The bigger the vig, the more the raw implied probability lies to you.

A Practical Rule

1. Convert both sides to implied probability.

2. Devig them before you compare to your model.

3. Only then decide if the gap is large enough to bet.

You cannot know if you have +EV until you know what "fair" even is.

Do you currently strip the vig before comparing a line to your number, or do you still compare against the posted odds?