What is the accuracy rate and return rate of horse racing AI? Calculation method and how to distinguish between expected results

The horse racing AI's accuracy rate shows how much you hit, and the return rate shows how much you got back compared to the purchase amount. Even if the hit rate is high, the payout amount may be less than the purchase amount, and even if the recovery rate is over 100%, there is a possibility that the result was exceeded by only a few high payouts. Please compare the performance of AI predictions by including the ticket type, total number of points purchased, purchase amount, refund amount, target period, and the time when the prediction was finalized.

This article is for those who want to know how to read and verify horse racing AI results. It is not intended to recommend any particular purchase or profit. Horse racing tickets are for people over 20 years old, and when purchasing them, please first decide on a reasonable budget. You can also check the age and purchase precautions in JRA horse racing ticket rules.

JRA horse racing ticket ruleshttps://www.jra.go.jp/kouza/baken/index.html

What is the difference between horse racing AI's accuracy rate and return rate?

Here, one point is defined as "one purchased item". If you buy 10 items and get 2 hits, the hit rate for each item you buy is 20%. The recovery rate is calculated from the total purchase amount, including hits and misses, and the total refund amount.

IndicatorCalculation formulaWhat can we learn
Hit rate per buy itemNumber of hit buy items ÷ Total number of items purchased × 100Percentage of purchased items that were hit
Recovery rateTotal refund amount ÷ Total purchase amount × 100Percentage of money returned relative to invested amount
Balance balanceTotal refund amount − Total purchase amountDifference from purchase amount. Does not include taxes

Be wary of achievements that just say "hit rate." There is also a "hit rate for each race" which is expressed as Number of hit races ÷ Number of races purchased, and if you buy multiple points for one race, the numbers will not match the number for each purchase. The 1st place rate and top 3 rate of predicted horses are also different from the accuracy rate of horse racing tickets. Check the numerator and denominator before comparing.

Why is the recovery rate low even if the hit rate is high?

The following is a hypothetical example to illustrate the calculation. In both cases, 10 items are purchased for 100 yen each, and the purchase price is 1,000 yen. This is not the performance of an actual AI.

Hypothetical predictionNumber of hit points/number of purchased pointsTotal refund amountHit rateCollection rateBalance balance
A4/10 points800 yen40%80%−200 yen
B2/10 points1,300 yen20%130%+300 yen

Even if A wins a lot of times, he cannot recover the purchase amount including the winnings. B will outperform the purchase price with these 10 points, but the next 10 points will not necessarily give the same result. If the ticket type and the odds of the selected horse are different, the relationship between the hit rate and the return rate will also change. You can check the winning conditions such as single win, double win, wide, triple, etc. under JRA horse racing ticket type.

JRA horse racing ticket typehttps://www.jra.go.jp/kouza/beginner/baken/index.html

Is the 100% recovery rate the same as JRA's refund rate?

A recovery rate of 100% means that the total purchase amount and refund amount are equal. For example, if you bought a total of 10,000 yen worth of horse racing tickets and received a total refund of 10,000 yen. In this case, there is no profit. The return rate here is the purchase result of individuals and expected services.

On the other hand, the refund rate determined by JRA for each ticket type is a systematic rate that calculates the amount to be refunded to winning bidders from the proceeds. This is not the same number as the individual recovery rate. You can check the set refund rate for each ticket type in JRA refund calculation method. When tabulating, please also record the treatment of purchases that were canceled or returned.

JRA refund calculation methodhttps://jra.jp/faq/pop03/1_17.html

What confirms the track record of "high return rate of horse racing AI"?

The published results of AI predictions include backtesting recalculated using past data,post-publication performance of predictions published before the race, andpurchase history of horse racing tickets actually purchased. It is important not to treat these three as the same achievement. Even if you look at Example of backtest using past races and Example that clearly shows how to calculate simulation recovery rate, the assumptions of aggregation affect the interpretation of the results. ZETTAI does not verify or guarantee the actual values ​​of linked sites.

Example of backtest using past raceshttps://bigbridge-keiba.jp/backtest

Example that clearly shows how to calculate simulation recovery ratehttps://analytics.keiba.link/archive/simulation-method/

Items to checkQuestions for readers
Target and denominatorEither central or local. How many races and points did you target? Does it include the missing part?
Ticket types and purchasing rulesWhat kind of ticket (single win, double win, etc.), how much per ticket, and under what conditions
Types of resultsEstimates based on past data, predictions fixed before the race, or actual purchases
PeriodAre there start and end dates, monthly fluctuations, and results for the most recent unused period
Basis for refundsWas the calculation based on the confirmed order of arrival and refund amount? How we handled restitution
Models and conditionsDid you record the version of the model, the time the forecast was issued, and any changes to the buying method?

It is not possible to recalculate the results for all purchases using only images that extract only the hits or the "highest collection rate." Prioritize reading the table that shows not only the percentage, but also the total purchase amount and refund amount, and the denominator points.

Even if the backtest is good, why can't it be reproduced in actual operation?

When performing verification using past data, be careful of data leaks where information that was not known at the expected time is mixed in. If you use the finish order after the race, the payout amount, and the final odds that were not determined at the time of the race as input for predictions, you will get results that cannot be reproduced before the actual race.

Also, if you look at the results for the same validation period many times and adjust the conditions, it may fit only that period. Use older periods for training, later periods for validation, and finally unused recent periods. It would be desirable to record the predictions made before the race with the date and time so that they can be checked against the final results later. If you want to know more about how to make it, please go to How to make horse racing AI and chronological verification.

How can I create a comparison table?

Regardless of whether or not you want to make a prediction or purchase, we will first decide on the target race and fix the prediction and purchase rules before the race. By recording each line of your purchases and aggregating them by month and ticket type, you can check whether you are relying on a small number of high dividends.

Columns to recordEntry example/purpose
Date, race, ticket typeSpecify comparison targets and winning conditions
Estimated release time/model versionCheck if you have selected it after viewing the results
Buy items, number of points, and purchase price per itemRecalculate the denominator of the hit rate and purchase price
Final result/refund amount/refundRecalculate collection rate and balance balance
Non-purchase races and reasonsMake sure you are not choosing only convenient races later

For example, even for the same 10 races, if you buy 1 point each time or 10 points each time, the way you read the winning rate for each race and the winning rate for each buy will be different. JRA's How to view refunds after confirmation can also be used for verification.

How to view refunds after confirmationhttps://www.jra.go.jp/kouza/baken/index.html

FAQ

Is the AI ​​easier to win if it has a high hit rate?

You can't tell by just the hit rate. Please check the number of purchases, ticket type, odds, purchase amount, and refund amount.

If AI has a recovery rate of over 100%, will it continue to be profitable?

This cannot be guaranteed. If the target period and number of purchases are small, a single high dividend may increase significantly. Please check the trends by period and the results of all cases after publication.

Are predicted win rate and hit rate the same?

it's different. The predicted win rate is the probability given by AI, and the hit rate is the rate calculated from the results of the actual purchases. Please read separately what was predicted and what was the denominator.

References and next article to read

Document confirmation date: November 12, 2026.

The related creation steps are explained in How to create horse racing AI | From data preparation to predictive model verification.