I Let the Model Ruin My Betting Card

The noble plan, the three bets I liked, and the model that disagreed with all of them #

I opened the workflow with three bets I liked and the polite hope that the model would admire my taste. The first bet was a road favorite I had been writing about all week; the second was an over on what felt like a fun primetime game; the third was a teaser with two key numbers I had decided constituted research. The model returned a score for each. Two of the three sat in the red. The third was tagged neutral, which was the model s polite way of telling me the bet was a coin flip with a hold attached.

My first instinct was to assume the model was confused. I checked the inputs. The inputs were correct. I checked the version. The version was current. I made a coffee and came back to the screen, hoping the numbers had reconsidered. They had not. This was the first useful lesson of the experiment, which I did not yet recognize as a lesson because it felt like an irritant. A tool that you ask to confirm you is a tool you have downgraded into a co-worker. The point of building a tool is to have a colleague who will occasionally tell you no.

The mechanic of friction, and why it works #

The model is not a wise teacher. It is a friction device. Most of what it does, when integrated into a betting workflow, is slow the bettor down at the moment of placement. It asks for an explicit input where the bettor would otherwise have used a vibe. It prints a number where the bettor would otherwise have used a feeling. It shows a probability where the bettor would otherwise have used a story. None of these substitutions is magic. They are the small inconveniences that prevent the bettor from clicking through her own gut while it is still warm.

The point of the friction is not that the model is always right. It is that the bettor is always interruptable. The first useful output of any model, in any field, is the interruption itself. The interruption produces a sentence — what does the model see that I do not? — and the sentence produces a small audit. Most of the bets I passed on, after the audit, I do not regret passing on. Some of the bets I took anyway, after the audit, I do not regret taking. The audit is the product. The pass or the placement is the result.

What the disagreement log recorded that my memory would not have #

I kept the disagreement log in a single tab with six columns: date, bet, my reasoning, model output, response taken, and a result column updated at settlement. The first observation, after about three weeks, was that the model was right considerably more often than my pre-experiment intuition had estimated. I had assumed the disagreements would split roughly fifty-fifty. The actual split, when I cleaned the log at the end of the year, was closer to thirty-one to fourteen in the model s favor on the disagreements where one of us was clearly wrong. The other thirty-eight disagreements were noise or shared errors, which is itself a lesson — most disagreements are not the binary I had imagined.

The second observation was about my response strategy. The disagreements I passed on returned a small positive ROI, on average. The disagreements I took my side on anyway returned a small negative ROI. The disagreements I sized down on were essentially flat. The disagreements I inverted — took the model s side instead of my own — returned the highest ROI of any response, but those were rare enough that I do not consider the number statistically meaningful. The robust finding was simpler: when the model and I disagreed, the pass was usually the right action, and the click was usually the wrong one.

The bets I would have placed without the friction #

The bets the model talked me out of were not bad bets in the spectacular sense. They were bets at slightly worse prices than I would have admitted at placement, on theses I had constructed with more affection than research. The road favorite from week three was a team I had spent the week defending in the group chat, and the price had drifted two points toward the favorite while I was busy defending it. The over on the primetime game depended on a pace assumption I had not actually checked. The teaser was, in retrospect, a teaser, which is a category of bet I tend to enjoy more than the bankroll does.

None of these failures were egregious. All of them were small. The point of the workflow is not that any one of them was decisive. The point is that the workflow caught about two-thirds of them across the season, and the bets I avoided produced, in aggregate, a meaningful improvement in the per-bet ROI. The improvement was not the model finding edges I had missed. It was the model letting fewer of my own marginal bets slip through the placement step.

The useful thing about a tool is not that it makes the decision for you. The useful thing is that it interrupts the part of your brain already writing the celebration tweet.

The discipline I want to keep for next season #

I have decided to keep the disagreement log as a standing feature of my workflow. The discipline is small: every bet I am about to place runs through the model first. If the model agrees, I place the bet at the originally intended stake. If the model disagrees, I write the disagreement into the log with a single sentence on my reasoning and a single sentence on what I read of the model s reasoning. The action defaults to a pass; placing the bet anyway requires me to write a third sentence explaining why my reasoning specifically beats the model s on this case.

This is not heroic. It is paperwork. It is also the only thing I have done in two years of betting that I can confidently attribute to a measurable improvement in my per-bet ROI. The improvement is modest in any single week and meaningful across a season. The cost is about thirty seconds per bet and the occasional bruised feeling when the model declines to admire a thesis I am proud of. The bruised feeling has, over time, become useful information in its own right. The bets I most resent the model for questioning have turned out, with frustrating regularity, to be the bets that most deserved questioning.

The closing argument #

The model did not make me a sharper handicapper. It made me a slower one. The slowness produced a smaller card of bets per week and a higher per-bet ROI on the bets that survived the workflow. None of this is the kind of result you can quote at a tailgate. It is the kind of result you can defend in a spreadsheet, which is a different but more durable form of correctness. The bets I most enjoyed talking about were, with painful regularity, the bets the model most wanted me to pass on. The bets the model approved of were, with reassuring regularity, the bets I could still defend in March.

I will keep the workflow. I will keep the log. I will keep the three-sentence rule for overriding the model, and I will keep being mildly annoyed when the rule prevents me from clicking on a bet I am sentimentally attached to. The annoyance is the feature. The bet I most regret not placing, across the entire season, lost cleanly by the way, which is the kind of detail the workflow makes available to me and which the version of me without the workflow would have spent the whole season insisting was the one that got away.