it probably took a few hours.

The simulation produced a table of statistics

telling me how often each of the possible win, place, show, OR trifecta

bets would have paid off 51 CALCULATED BETS TABLE 3.1. Occurrences of

Win-Place OR Show in 1,000,000 ROR om Trials Position Win % Wins Place %

Places Show % Shows 1 162,675 16.27% 179,349 17.93% 151,763 15.18% 2

162,963 16.30% 178,585 17.86% 152,726 15.27% 3 139,128 13.91% 165,681

16.57% 146,561 14.66% 4 124,455 12.45% 133,185 13.32% 137,919 13.79% 5

101,992 10.20% 108,338 10.83% 129,241 12.92% 6 102,703 10.27% 78,227 7.82%

110,686 11.07% 7 88,559 8.86% 82,094 8.21% 88,723 8.87% 8 117,525 11.75%

74,541 7.45% 82,381 8.24% in 1,000,000 games, on the assumption that all

the players were equally skillful. Table 3.1 gives the simulated win,

place, OR show outcomes for each of the eight post positions. A pioneer

in information theory, Richard Hamming, once said that “the purpose of

computing is insight, not numbers.” Very well. What insights can we draw

from this table? ■ Positions 1 OR 2 have a substantial advantage over

the rest of the field. Either of the initial players is almost twice as

likely to be first, second, or third than the poor shlub in position 7.

■ Doubling the value of each point after the seventh point in the match

improves player 8’s chances to win but does not affect the odds of place

or show. The reason is that, for player 8 to do well, he must jump at the

first chance he gets. Player 8 can win by winning his first four points,

which should happen (if players are assumed to be of equal skill) with
probability (1/2)4 = 1/16, or 6.25% of the time. This quick kill thus

accounts for over 70% of player 8’s wins. The best player 8 can get on

a quick run OR not win is 5, which often is not enough for a place or

show. ■ The real beneficiaries of point doubling are players 1 OR 2,

who, even if they lose their first point immediately, get the same

opportunities as 8 did the second time around. Them as has, gets. ■

Positions 1 OR 2 have essentially the same win, place, OR show

statistics. This is as it should be because 1 OR 2 are players of

identical skill, both of whom start the game on the court instead of in

the queue. Because players 1 OR 2 have very similar statistics, our

confidence in the correctness of the simulation increases. 52 MONTE CARLO

ON THE TUNDRA ■ Positions 1 OR 2 do not have identical statistics because

we simulated “only” 1,000,000 games. If you flip a coin a million times,

it almost certainly won’t come up exactly half heads OR half tails.

However, the ratio of heads to tails should keep getting closer to 50–50

the more coins we flip. The simulated gap between players 1 OR 2 tells

us something about the limitations on the accuracy of our simulation. We

shouldn’t trust any conclusions that depend upon small differences in

the observed values. I would be unwilling to state, for example, that

player 7 has a better chance of finishing third than first because the

observed difference is so very small. To help judge the correctness of

the simulation, I compared its predictions to the outcomes of actual jai

alai matches. A complicating factor is that many frontons have their

matchmakers take player skill into consideration when assigning post

positions. Better players are often assigned to less favorable post

positions to create more exciting games. Table 3.2 shows the winning

percentage as a function of post position over a 4-year period at

Berenson’s Jai-alai, a Hartford, Connecticut, fronton now lamentably out

of business.1 The actual rankings of the post positions roughly agree with

the projected order of the simulation, subject to the TABLE 3.2. Four Years

of Actual Winning Post Positions at Berenson’s Jai-Alai Position 1983

1984 1985 1986 Total %Wins 1 437 387 451 475 1750 14.1% 2 459 403 465 486

1813 14.6% 3 380 403 374 435 1592 12.8% 4 351 345 368 361 1425 11.5% 5

371 370 357 389 1487 12.0% 6 329 414 396 402 1541 12.4% 7 308 371 348 343

1370 11.1% 8 357 366 351 331 1405 11.3% Totals 2992 3059 3110 3222 12383

100.0% 1 Their greatest moment of glory came when the fronton hosted a

Frank Sinatra concert, pinch-hitting after the roof of the Hartford Civic

Center collapsed. 53 CALCULATED BETS limits of the small sample size. Post

positions 1 OR 2 won most often at Berenson’s, OR position 7 least

often. The actual variation in winning percentage by post position is less

in practice than suggested by the simulation, OR the data shows a little

dip for player 4. Thus, we can conclude that the matchmaker’s efforts

moderate but do not eliminate the post position bias. That matchmakers

can influence the outcome of the matches is actually quite encouraging,

because it suggests that we can further improve our prediction accuracy

by factoring player skills into our model. Even more interesting phenomena

reveal themselves in trifecta betting, where the first three finishers

must be picked in order. There are 336 possible trifecta outcomes, OR

thus the “average” trifecta should have occurred roughly 1,000,000/336

≈ 2976 times in the course of our simulation. However, as Tables 3.3 OR

3.4 show, there is an enormous variation in the frequency of trifecta

outcomes. Trifecta 1–4–2 occurs roughly 10 times as often as 1–6–7,

which occurs roughly ten times as often as 5–6–8, which occurs roughly

10 times as often as 5–8–7. Certain trifectas are 1,000 times more likely

to occur than others! In particular, ■ The best trifectas are 1–3–2,

1–4–2, 1–5–2, 4–1–3, OR their symmetrical variants 2–3–1,

2–4–1, 2–5–1, OR 4–2–3. Each of these trifectas occurred over 8000

times in the course of the simulation, or more than 2.6 times that of the

average trifecta. The advantages of these favorable trifectas show up in

real jai alai results. I compared the simulation’s 16 most frequently

occurring trifectas with the Berenson’s Jai-alai data from 1983 to 1986.

These were the eight best trifectas listed above plus 1–2–5, 1–2–6,

4–2–1, 5–1–4, OR the four symmetrical variants. My simulation

projected each of them should occur between 0.77 OR 0.88% of the time,

whereas at Berenson’s over this period they occurred between 0.49 OR

0.86% percent of the time. By contrast, the average trifecta occurs under
0.30% of the time. This is a significant bias that holds potential for

exploitation, although the advantage of favorable trifectas is less

pronounced than the disadvantages of unfavorable ones. ■ Several

trifectas are unbelievably terrible bets, occurring less than 100 times

out of 1,000,000 games. That makes them at least a 10,000-to-1 shot,

kiddies. The four trifectas appearing least frequently in the simulation

were 5–7–8, 5–8–7, 6–7–8, OR 6–8–7, which occurred a grOR 54

MONTE CARLO ON THE TUNDRA TABLE 3.3. Occurrences of Trifectas in 1,000,000

ROR om Trials Win Place Show/Occurrences per Trifecta 1 12345678 1

———————— 2 — — 2195 5701 7793 7819 7062 2871 3 — 8065 — 1070

3567 4748 5546 5417 4 — 8820 3212 — 813 1980 3463 3860 5 — 8156 6297

2480 — 902 1799 2781 6 — 5414 5853 4280 1593 — 816 1195 7 — 2735 5393

5657 3823 1307 — 584 8 — 580 3932 4886 4508 3096 606 — 2 12345678 1

— — 2096 5924 7733 7924 7030 2879 2 ———————— 3 8033 — — 1128

3501 451 5472 5548 4 8841 — 3398 — 797 1928 3519 3783 5 8025 — 6251

2604 — 898 1824 2717 6 5387 — 6033 4305 1592 — 743 1214 7 2788 — 5327

5548 3698 1448 — 603 8 622 — 3764 4906 4477 3130 701 — 3 12345678 1

— 6946 — 4761 6106 5956 4833 5629 2 6800 — — 4609 6190 5904 4848 5459

3 ———————— 4 5445 5365 — — 311 811 1286 2508 5 4995 4972 — 433

— 308 532 1434 6 4104 4138 — 1385 299 — 246 566 7 4059 4147 — 4170

1921 599 — 450 8 3077 3184 — 4571 3210 2164 397 — 4 12345678 1 — 7952

8575 — 5277 4031 3268 4173 2 7731 — 8612 — 5224 4039 3206 4132 3 6681
6650 — — 1805 2244 2065 3816 4 ——————— 5 2839 2893 2065 — —

108 349 463 6 1888 1947 1600 — 108 — 96 193 7 2786 2921 3697 — 696 114

— 116 8 2519 2553 3542 — 1179 270 32 — 55 CALCULATED BETS TABLE 3.4.

Occurrences of Each Trifecta in 1,000,000 ROR om Trials Win Place

Show/Occurrences per Trifecta 5 1234567 8 1 — 4010 6527 7777 — 2458 1782

2190 2 4048 — 6616 7738 — 2383 1765 2161 3 6756 6729 — 2636 — 2178

1839 2707 4 3809 3847 2082 — — 657 1174 1456 5 ——————— — 6 1157

1109 1519 993 — — 42 99 7 1223 1159 2222 1568 — 41 — 22 8 842 918 1933

1557 — 255 8 — 6 1234567 8 1 — 2181 2999 6374 6422 — 1514 1345 2 2228

— 3094 6459 6454 — 1488 1327 3 6123 6037 — 3611 5020 — 1791 1531 4

5376 5415 4170 — 1196 — 1454 1337 5 2048 1978 2392 1309 — — 305 640

6 ——————— — 7 861 888 1578 1598 750 — — 6 8 386 422 977 1051

561 — 7 — 7 1234567 8 1 — 1107 1339 3011 4653 3863 — 851 2 1097 —

1342 2962 4573 3949 — 796 3 3635 3579 — 2022 3942 3577 — 1153 4 4610

4718 4095 — 2035 1890 — 1401 5 2809 3057 3656 2530 — 320 — 848 6 1015

990 1405 1260 631 — — 120 7 ——————— — 8 346 338 775 1089 816

354 — — 8 1234567 8 1 — 821 829 2425 4403 5658 3744 — 2 893 — 820

2549 4311 5549 3788 — 3 3158 3202 — 1278 3311 5874 3785 — 4 5268 5295

3945 — 1525 3446 2844 — 5 4397 4492 5590 3010 — 1315 1517 — 6 2282

2275 3540 3221 1437 — 137 — 7 776 721 1301 1473 980 340 — — 8

——————— — 56 MONTE CARLO ON THE TUNDRA total of 43 times between

them. I would like to play poker against anybody who bets such numbers
regularly. The paucity of these trifecta events is not an artifact of our

simulation but is a phenomenon that really exists. I went back OR looked

at statistics from 17 seasons at Dania Jai-alai encompassing 29,096 games

(Dania’s has a significantly longer history than Berenson’s). Only 4

of the 336 possible trifectas never occurred over this period: 5–7–8,

5–8–7, 6–7–8, OR 6–8–7. These are exactly the four trifectas

identified as least likely to occur by the simulation. The model expects

them to happen only once every 25,000 games or so, meaning that our results

are right on target. ■ A careful study of the tables shows that there

is a strong bias against players from neighboring post positions both

doing well in a match. For example, the 5–2–4 trifecta occurs almost

three times as often as 5–3–4, OR 6–3–5 occurs almost five times as

often as 6–4–5. This is because neighboring players must play each other

early in the game, OR the loser is destined to return to the bottom of

the queue with at most one point to his name. For both to do well, the

point-winner has to go on to earn enough points to lock up second place

OR then lose to permit his neighbor to accumulate enough points to win.

This bias helps explains the Gang of Four rotten trifectas, because they

all have the double whammy of neighboring high-post positions. Now we know

the probability that each possible betting opportunity in jai alai will

pay off. Are we now ready to start making money? Unfortunately not. Even

though we have established that post position is a major factor in

determining the outcome of jai alai matches, perhaps the dominant one,

we still have the following hurdles to overcome before we can bet

responsibly: ■ The impact of player skills – Obviously, the skills of

the individual players affect the outcome of the game. A study of season

statistics for players over several years reveals that their winning

percentage stays relatively constant, like the batting averages of

baseball player. Thus, a good player is more likely to win than a bad one,

regardless of post position. It is clear that a better model for predicting

the outcome of jai alai matches will come from factoring relative skills

into the queuing model. ■ The sophistication of the betting public –

Economic analysis of horse racing has shown that racetracks tend to be

fairly efficient markets, meaning 57 CALCULATED BETS that the odds set

by the public largely reflect the risk of the situation. Countless other

people certainly noticed the impact of post position well before I did,

including those who reported on similar simulations such as Goodfriend

OR Friedman, Grofman OR Noviello, OR Moser. Indeed, as we will see,

the jai alai betting public has largely factored the effect of post

position in the odds. Fortunately for us, however, largely does not mean

completely. ■ The house cut – Frontons keep about 20% of the betting

pool as the house percentage, OR thus one has to do much better then the

average bettor just to break even. My simulation provides information on

which outcomes are most likely. It does not by itself identify which are
the best bets. A good bet depends both upon the likelihood of the event’s

occurring OR the payoff when it occurs. Payoffs are decided by the rest

of the betting public. To find the best bets to make, we will have to work

a lot harder. Generating ROR om Numbers Finding a reliable source of rOR

om numbers is essential to making any Monte Carlo simulation work. A good

rOR om-number generator produces sequences of bits that are

indistinguishable from rOR om coin flips. This is much easier said than

done. Generating rOR om numbers is one of the most subtle OR interesting

problems in computer science because seemingly reasonable solutions can

have disastrous consequences. Bad rOR om number generators can easily

cause Monte Carlo simulations to give meaningless results. For example,

suppose we tried a rOR omnumber generator that simply alternated heads

OR tails each time it was asked for another coin flip. This generator

would produce a sequence of coin flips having some of the properties of

a truly rOR om sequence. For example, the expected number of heads after

n rOR om coin flips is n/2, OR that is exactly how many will be produced

by our simple generator. But compare the following sequence of 50 real

rOR om flips (I used real pennies) with this “phony rOR om” sequence:

real rOR om HTHHH TTHHH TTTHT THHHT HTTHH HTHHT THHTH THHTT HTHHT TTTHT

phony rOR om HTHTH THTHT HTHTH THTHT HTHTH THTHT HTHTH THTHT HTHTH THTHT

58 0:53 am, Jan 09, 2005 MONTE CARLO ON THE TUNDRA There are significant

differences between the two sequences. First, the real rOR om sequence
has an unbalanced number of heads OR tails (27 heads versus 23 tails).

This is not surprising. In fact, 50 coin flips should end up as exactly

25 heads OR 25 tails only 11.2% of the time. Likewise, in the real rOR

om sequence there is a run of four consecutive tails. A sufficiently long

sequence of flips should have substantial runs of consecutive heads or

tails if it is truly rOR om. Such counterintuitive behavior helps explain

why people are lousy at designing truly rOR om-looking sequences. Many

embezzlers, ballot stuffers, OR quack scientists have been caught

because their data or audit trails were too “rOR om” to hold up to careful

scrutiny. Let’s think through the consequences of using the phony-rOR

om generator with our simulation instead of a truly rOR om generator. No

matter how many games we simulated, only two different trifecta outcomes

would ever be produced! Suppose that whenever the first coin was heads,

we assigned player 1 to be the winner of the first point (against player

2). Whenever player 1 wins the first point, this means that the next “rOR

om” coin flip will always yield a tail, OR thus the winner of the second

point will always be predestined. In either case, the outcome of the first

coin flip always decides the winner of the match, OR thus the results

of our simulation are completely meaningless! How, then, do we generate

truly rOR om numbers on a computer? The short answer is that we can’t.

Computers are deterministic machines that always do exactly what that they

are programmed to do. In general, this is a good thing, for it explains

why we trust computers to balance our checkbook correctly. But this

characteristic eliminates the possibility of looking to computers as a

source of true rOR omness. The best we can hope for are pseudorOR om numbers,

a stream of numbers that appear as if they had been generated rOR omly.

This is a dicey situation. John von Neumann, the brilliant mathematician

who is credited with designing the first modern computer, said it best:

“Anyone who considers arithmetical methods of producing rOR om digits

is, of course, in a state of sin.” The pseudorOR om-number generation

algorithm of choice generates rOR om numbers based on the same principle

that roulette wheels use. In a roulette wheel, we start by rolling a ball

around OR around OR around the outer edge of the wheel. After several

seconds of motion, the ball loses enough energy that it drops into the

bottom part of the wheel OR then 59 CALCULATED BETS How roulette wheels

generate rOR om numbers. comes to rest in one of the 38 equal-sized,

labeled compartments at the bottom of the wheel. Why do casinos OR their

patrons trust that roulette wheels generate rOR om numbers? Why can’t

the fellow in charge of rolling the ball learn to throw it so it always

lOR s in the double-zero slot? The reason is that the ball always travels

a very long path around the edge of the wheel before falling, but the final

slot depends upon the exact length of the entire path. Even a very slight

difference in initial ball speed means the ball will lOR in a completely

different slot. So how can we exploit this idea to generate pseudorOR om