How a Tiny Information Edge Beats Market Smarts in Hong Kong Horse Racing

Understanding the System

Hong Kong horse‑racing uses a pari‑mutuel pool: all bets on a horse go into a single pot, the pool is closed before the race, a fixed takeout (about 19% in Bill Benter’s era) is removed, and the remaining amount is divided among winning tickets. Consequently, odds are determined by the distribution of bets, and the market functions as a negative‑sum game—on average each HK$1 wager returns only about HK$0.80 after takeout.

The 2025/26 season saw total betting turnover of approximately HK$143.3 billion, a 3.2% year‑over‑year increase, while the Hong Kong Jockey Club paid HK$28.8 billion in gambling and profit taxes, underscoring the market’s massive scale.

Can Anyone Win Long Term?

Bill Benter, an American mathematician, partnered with Australian gambler Alan Woods and built a five‑person‑year database and model that achieved profit in four of five seasons, losing about 20% of capital in the single losing year. Bloomberg Businessweek (Chellel, 2018) reported that Benter’s cumulative winnings approached US$1 billion.

Before celebrating Benter’s success, the article asks what level of competition he faced. Studies by Busche and Hall (1988) and Busche (1994) examined 2,000+ Hong Kong races and found no systematic “favorite‑longshot bias” that exists in some North American markets; public odds were generally well‑calibrated, with only extreme longshots (<1% win probability) being over‑estimated.

The Model

Benter employed the multinomial logit model introduced by Bolton and Chapman (1986). Each horse receives a score based on factors such as recent performance, weight carried, jockey, gate position, and distance suitability, which is transformed into a win probability that automatically sums to 1 across all horses in a race.

Because the model is incomplete—omitting trainer intent, horse condition on the day, and unpublished trial data—Benter discovered systematic error when directly comparing model estimates to actual outcomes. He split the data into two halves: races where public odds exceeded the model and races where the model exceeded public odds. In both halves, the actual win frequency leaned toward the public side, indicating that the market incorporated information the model missed.

To capture this, Benter treated the model estimate and the public odds as two expert opinions and performed a second‑stage logit fusion. The fusion weight for the public odds reflects its implied probability, while the weight for the model is derived from maximum‑likelihood estimation on historical data.

The goodness‑of‑fit is measured by pseudo‑log‑likelihood (the log of the likelihood under a naïve “all horses equally likely” assumption). Higher values indicate stronger explanatory power.

Using 3,198 races from 1986‑1993, Benter reported three pseudo‑log‑likelihood values for public odds, the model, and the fused estimate. The model alone improved over public odds by only 0.0027, but the fusion added a meaningful increment.

A Comparative Test

Benter also evaluated a second sample (2,313 races from Sep 1988 – Jun 1993) with a simplified nine‑factor model and a “tip‑expert” estimate derived from weighting newspaper tipsters’ top‑three predictions (6 points for first, 3 for second, 1 for third, then normalized).

Individually, the public odds, nine‑factor model, and tip‑expert each explained a portion of the outcome, but both the model and tip‑expert performed worse than the market overall. After fusing each with public odds, the nine‑factor model contributed an additional 0.0090 to pseudo‑log‑likelihood, whereas the tip‑expert added only 0.0002, showing that marginal explanatory power does not correlate directly with standalone accuracy.

The tip‑expert’s negligible contribution is explained by the fact that newspaper predictions are already reflected in the odds; when tip‑expert and odds diverge, the odds are usually correct.

Benter’s profit stemmed from the portion of his error that was uncorrelated with the market’s error—e.g., a distance‑preference factor derived from multivariate regression that the public never calculated. The value lies in information that is new, not merely more accurate.

Key Takeaways

The logic extends beyond horse racing: wherever a collective price exists (stocks, salaries, etc.), personal judgments must first subtract the consensus component. An insight that is already priced yields little extra return.

Research using Hong Kong odds from 1979‑2023 shows the public’s pseudo‑log‑likelihood has risen over four decades (0.1325 for 1986‑1993, 0.1437 for 1996‑2003, 0.1668 for 2006‑2013, 0.1863 for 2016‑2023), indicating the market has become increasingly efficient as model‑driven bettors like Benter inject information that is then absorbed into the odds.

Benter’s 1994 paper concludes that the era of profitable computer‑based handicapping is fleeting; as more analysts publish models, the market will price them in, and only the earliest, well‑engineered models will capture excess returns.

Ultimately, the value of a piece of information depends on how many others already know it and how much time remains to exploit it.