Odds Compilers Are Being Replaced by Algorithms - What Does That Mean for Finding Edges?

[SharpEddie47]

The transition has been happening for years but I want to talk about what it actually means practically.

In 2005 the opening price on a Cowboys game was set by a human being. Someone who knew the NFL, had opinions, made judgment calls, and occasionally made systematic errors you could identify and exploit.

The human compiler who consistently underrated road underdogs in divisional games. The one who overweighted recent form relative to schedule strength. The one who hadn't fully processed an injury report by opening time on Tuesday.

These were real edges. Documented in my records. Exploitable repeatedly until they closed.

They closed because those humans were replaced by models.

The models don't have the same systematic errors. They have different ones.

The question is whether the algorithmic blind spots are identifiable and exploitable the same way human blind spots were.

My working hypothesis after five years of trying: the algorithmic errors exist but they're harder to find, less persistent when found, and correct faster when exploited.

What's everyone else seeing.

 

[FadeThePublic]

The public money pricing question is the one relevant to my approach.

Human compilers priced public sentiment inconsistently. Some were aggressive about shading lines toward public action. Some weren't.

You could develop a map of which operators adjusted aggressively for public money and which didn't.

The gaps between operators who shaded and operators who didn't created cross-book arbitrage opportunities around public-heavy games.

Algorithms have largely eliminated this inconsistency.

Major operators now use similar algorithmic approaches to pricing public sentiment. The gaps between them have compressed significantly.

The cross-book inconsistency that was exploitable in 2010-2015 is mostly gone.

What remains: the algorithm's model of public sentiment is built on historical data. When public behavior changes in ways the model hasn't seen before, the pricing can lag.

The new sport for fading the public is finding moments where public behavior is genuinely novel and the algorithm's historical training is stale.

Those moments exist. They're rarer and shorter-lived than they used to be.

 

[Klaus]

The Bundesliga algorithmic transition has a specific timeline I can track against my model's performance.

2015-2018: meaningful edge identifiable in opening prices. Human compiler systematic errors visible in specific market types.

2019-2021: rapid efficiency increase. The operator I primarily used for opening price value switched to predominantly algorithmic compilation in this period.

Post-2021: the opening price value largely gone. The edge in my approach shifted from exploiting opening price errors to identifying closing line divergence from my model.

The human compiler errors were systematic and learnable.

The algorithmic errors are different. They exist but they're statistical rather than systematic. Emerging from specific training data gaps rather than human cognitive patterns.

The algorithmic blind spots require different methodology to identify.

I'm still building that methodology.

 

[oli_sussex]

The exchange is an interesting case because it was never human-compiled.

Betfair's prices emerge from the interaction of participants. The "compilation" is the market itself.

In principle this should produce the most efficient prices available because it aggregates all participant information.

In practice the exchange has specific inefficiencies that differ from both human-compiled and algorithmic prices.

Late market information isn't always incorporated quickly because it depends on participants acting on it.

Thinly traded events have wide spreads because there aren't enough participants.

Algorithmic compilation covers events the exchange can't cover efficiently because participation is insufficient.

The exchange, human compilation, and algorithmic compilation each have different failure modes.

The sophisticated bettor maps which failure mode applies to which market and targets accordingly.

 

[TaffyTipster]

Not sophisticated enough to have tracked this directly.

But I've noticed something at the level of rugby prices.

The Premiership and Six Nations markets: prices feel tighter and move faster than they used to.

Matches involving smaller Welsh clubs in the Challenge Cup or lower-tier European competition: prices still feel soft. Still move in ways that suggest less computational attention.

The top markets got the algorithm. The bottom markets still feel like someone less attentive is compiling them.

Whether that someone is a human or a less sophisticated algorithm I don't know.

But the quality differential between top and bottom is visible even to someone who doesn't track it systematically.

 

[ParlayPrincess_88]

I genuinely didn't know what an odds compiler was before this thread.

I assumed the prices just came from somewhere automatically. Like a computer somewhere calculating it.

Learning that it was humans with specific knowledge and then those humans were replaced by actual computers.

There were people whose entire job was setting the prices on sports events.

And then there weren't.

Or there are fewer of them.

That's a strange thing to think about as someone who places bets without thinking about where the price comes from.

 

[CoachTony_Bets]

The coaching information angle has a specific relationship to this transition.

A human compiler who knew football deeply might adjust for soft information.

Locker room chemistry. A coordinator's tendencies in specific situations. A head coach's behavior under pressure.

This stuff isn't in any dataset. It's in conversations between people who know the game.

Algorithms are trained on quantifiable historical data. They're structurally blind to qualitative information that hasn't been converted to data.

The algorithmic transition may have actually increased the value of soft information for bettors who have legitimate access to it.

Not inside information in the problematic sense.

Just deep contextual understanding that hasn't been captured in a training dataset.

 

[FadeThePublic]

Tony's soft information point is interesting.

The human compiler's edge over the algorithm is exactly what Tony describes.

Qualitative judgment. Experience-based pattern recognition that predates quantification.

Those compilers are largely gone from tier one markets.

They still exist in tier two and three markets where algorithmic coverage is thinner.

The edge following the human compiler into lower tiers has some validity.

The question is whether the stake limits in those markets allow meaningful action.

 

[SharpEddie47]

The stake limits in algorithmically undersupported markets is the practical constraint that makes the theory only partially actionable.

Genuine inefficiency in a League Two match. Real edge identified.

Maximum stake: £50 at most operators. Often less.

The edge exists. The opportunity to size it appropriately doesn't.

The algorithm concentrating computational resources on high-value markets means the inefficiencies in low-value markets are real but thin in both price and liquidity.

Finding the edge isn't the constraint. Capitalizing on it is.

 

[Klaus]

The liquidity constraint Eddie describes is the fundamental limit on the retreating-to-lower-markets strategy.

The algorithmic frontier moves downward over time.

Markets that were soft in 2015 because algorithmic attention was limited to top leagues now have algorithmic coverage because the cost of scaling models downward has decreased.

The inefficiency available in the Championship in 2015 is substantially reduced in 2024.

The inefficiency available in League One in 2015 is substantially reduced.

The frontier keeps moving down.

The retail bettor following the frontier finds themselves in increasingly illiquid markets with decreasing ability to size positions appropriately.

The retreat is valid as a direction of travel.

The destination is markets too thin to matter.

 

[oli_sussex]

The frontier movement Klaus describes has a name in market microstructure theory.

Algorithmic attention is a resource that gets allocated based on economic value.

High-value markets attract maximum algorithmic attention earliest.

Low-value markets attract it later as the cost of extending coverage falls.

The retail bettor's retreat into lower markets is a race against the decreasing cost of algorithmic coverage.

The bettor can stay ahead of the frontier only as long as they can find markets where the economic value is insufficient to justify algorithmic deployment.

Those markets get harder to find as computing costs fall.

 

[DublinDegen]

not relevant to how i bet...

but reading this thread thinking about what happens if algorithms really do eventually price everything efficiently...

what does betting become then...

if there's no edge anywhere for anyone except institutional operators with proprietary models...

does the recreational market just become pure entertainment product with no pretense of skill...

 

[TaffyTipster]

Conor that's probably the destination.

The mass market becomes what the SGP thread and the micro-betting thread described.

Entertainment product. Acknowledged house edge. Sold on engagement not returns.

The genuine edge space becomes exclusively institutional or not at all.

 

[FadeThePublic]

The recreational market has arguably already arrived at that destination for most participants.

The pretense of skill is currently doing a lot of work in the industry's self-presentation.

"Sports betting requires skill and knowledge."

True for a small number of participants. Marketing for everyone else.

The algorithmic transition accelerates the hollowing out of the skill claim at the retail level.

 

[ThePuntingProf]

I started betting seriously when human compilers were the only compilers.

The specific edges from that period were learnable and persistent because the compiler's individual characteristics were stable.

You could study a market the way you study an opponent. Learn their tendencies. Exploit the gaps.

The algorithm has no tendencies in that sense. It has training data gaps and model assumptions that are harder to characterize as tendencies.

What I've found after thirty years: the algorithmic era requires different skills.

Less about learning an opponent's habits.

More about understanding what data a model was trained on and what situations fall outside that training distribution.

The second skill requires knowing more about the model than the model's creators typically disclose.

That's a different kind of research than I did for the first twenty years.

I'm still developing it at fifty-eight.

Whether I have time to fully develop it before the frontier passes me is a question I don't examine too carefully.

 

[SharpEddie47]

Prof naming the specific skill transfer required.

From: understanding human systematic error patterns.
To: understanding algorithmic training distribution gaps.

These require different expertise and different research methodologies.

The bettor who was expert at the first is not automatically expert at the second.

There's a retraining cost that nobody in the industry has been honest about.

The edge didn't just get smaller. It changed shape.

And the people whose edge was shaped for the human compiler era are navigating something that requires rebuilding from a different starting point.

 

[Klaus]

The retraining cost is the thing I'm currently experiencing.

The methodology that produced strong results for a decade is producing diminishing returns.

The methodological adjustment required is not a parameter tweak.

It's a structural change in what I'm looking for and how I'm looking for it.

Fourteen years of expertise in human-era edge identification.

Beginning the process of developing expertise in algorithmic-era edge identification.

The timeline for that development is unclear.

The market will not wait for me to complete it.

 

[oli_sussex]

The exchange model has its own version of this problem.

The participants who understood how human-compiled prices created opportunities for exchange trading had to relearn when algorithmic prices became the reference market.

The exchange price forms in relation to the algorithmic reference price.

Understanding algorithmic pricing became necessary to understand exchange dynamics.

One methodology built on another.

The layers of required knowledge deepened while the margins available compressed.

More expertise required for less return.

That's the trajectory the shrinkflation thread described and this thread confirms.