記事の図解

A model needs an advantage over the price

Estimated probability, market-implied probability and realized results are different measures. Odds can change, and a historical edge can disappear once tested on new data.

What this article covers

  • Why horse racing is negative-sum for bettors as a group.
  • Why expected value matters more than hit rate.
  • How to read market-implied probability from odds.
  • Why overfitting and bankroll management failures are common.
  • How to think about tax, terms-of-use, and addiction risks.

First, the conclusion

  • Market odds are prices that reflect the view of the betting market.
  • Expected value appears only when your estimated probability is higher than the market-implied probability.
  • Any edge must remain after takeout, taxes, model error, and losing-streak risk.

Why the horse racing market is negative-sum

JRA payout rates are set by bet type. As of June 2026, based on JRA's official information, win and place bets pay out at 80%, bracket quinella, quinella, and wide bets at 77.5%, exacta and trio bets at 75%, trifecta at 72.5%, and WIN5 at 70%.

In other words, the takeout rate is roughly 20% to 30%.

Suppose the betting market buys JPY 1,000,000 worth of tickets. If the payout rate is 75%, the payout pool is JPY 750,000. The remaining JPY 250,000 is taken out.

Bettor wagers: JPY 1,000,000
 ->
Takeout: JPY 250,000
 ->
Payout pool: JPY 750,000

Look at expected value, not hit rate

Expected return = estimated probability x odds
Expected ROI = estimated probability x odds - 1

Suppose a horse is priced at odds of 12.0, and your model estimates its probability at 10%.

Expected return = 10% x 12.0 = 120%
Expected ROI = 120% - 100% = +20%

By contrast, even if you believe a 1.1-odds horse has an 80% chance of coming in:

Expected return = 80% x 1.1 = 88%

Market-implied probability and estimated probability

Market-implied probability ~= 1 / odds

If the odds are 5.0, the market is roughly implying a probability of around 20%.

Estimated probability > market-implied probability

For example, the market may be pricing a horse as if its chance is around 20%, while your model estimates 25%. If that gap truly exists, there is a price distortion.

Overfitting and loss capacity

Kelly calculations depend on an accurate estimate that may not exist. Smaller stakes do not turn negative expected value positive. Tax treatment and operator rules also require independent examination.

Dynamic edge disappears quickly

  • A horse returning from a specific outside training facility.
  • The interaction between a bloodline and a particular track condition.
  • A horse that suffered serious traffic trouble in its previous race.
  • Conditions involving a jockey or stable that tends not to attract public money.
  • Inside or outside draw bias.
Discover a feature
 ->
Market money enters
 ->
Odds adjust
 ->
Expected value thins out
 ->
Search for the next distortion
  • Has recent ROI declined?
  • Are odds for a specific condition lower than they used to be?
  • Is expected ROI still there, rather than just hit rate?
  • Are you chasing a feature that the market has already priced in?

Overfitting is the biggest enemy

  • Running time
  • Track condition
  • Draw
  • Distance aptitude
  • Bloodline
  • Jockey
  • Trainer
  • Course configuration
  • Previous-race details
  • Race spacing
  • Odds movement
  • Split training and validation periods while preserving time order.
  • Check ROI in out-of-sample periods.
  • Do not overtrust tiny samples created by overly narrow race conditions.
  • Look at maximum drawdown, not only ROI.
  • Check hit rate, average payout, losing streaks, and the bankroll curve together.

Evaluate risk-adjusted returns, not just expected value

Item to checkMeaning
Average expected ROITheoretical edge per ticket
VarianceVolatility of returns
Maximum drawdownHow far bankroll can fall
Losing-streak probabilityWhether the strategy can be executed continuously
Odds movementWhether expected value remains at purchase time
After-tax returnWhat actually remains

Bankroll management: think fractional Kelly, not full Kelly

f* = (p x b - q) / b
SymbolMeaning
f*Fraction of bankroll to wager
pEstimated probability
qProbability of losing, or 1 - p
bOdds multiple minus 1
  • Reduce to half Kelly or quarter Kelly.
  • Set a maximum stake per race.
  • Set a daily or meet-level loss limit.
  • Do not increase stake size after losing streaks.
  • Never use money needed for daily life.

Tax risk can be heavier than model risk

The National Tax Agency indicates that payouts may fall under miscellaneous income in cases such as when a person uses original conditions or purchase patterns, buys many tickets throughout the year, and it is objectively clear that the overall return rate over the relevant period exceeded 100%.

Keep the claim proportionate to the evidence

Drawdowns, all stakes, out-of-sample results and costs are needed to assess a strategy. An attractive backtest is insufficient evidence of future profitability or permissible automated access.

Be careful with words like automated purchase and API

Feed it into an automated order system
Check the rules and terms of use, then maintain a purchase process and records.

Minimum metrics for horse-racing quant analysis

MetricWhy it matters
Hit rateFrequency of wins and losses
Average oddsPayout level when winning
ROIReturn relative to purchase amount
Expected ROIExpected profit per unit
Maximum drawdownHow much the bankroll can decline
Losing streak lengthPsychological and bankroll durability
Sample sizeDistinguishing luck from skill
After-tax returnAmount that actually remains

Checklist for a strategy note

Check the takeout rate
 ->
Read the market-implied probability
 ->
Estimate probability
 ->
Calculate expected value
 ->
Question overfitting
 ->
Set bankroll management rules
 ->
Check tax, terms-of-use, and addiction risks

Sources