You are on page 1of 17
Section Four Money Management The importance of money management and bet slz- ing has been stressed increasingly in recent years and rightly so. For even if the player has discovered a favorable betting situation, how he wagers determines. his success or failure. Ultimately itis the “bottom line” ‘on which a gambler’s performance Is judged. 1's tine, of course, to describe the favorable situation toa friend or business associate, but the next question is likely fo be “How much money are you making from this situation?” ‘The problem for the gambler is that much of the ad- vice on money management is conflicting or confus- Ing, or simply based on false premises. There are nun- dreds of schemes designed to overcome the house ‘edge in roulette and craps based solely on manipulating the size of one's bets. AS will ba saan, all such attempts are futile Even assuming the player has discovered a favorable {game (le, one offering a positive expectation), the ques- tion naturally arises: How does one best use a limited amount of capital to exploit this positive expectation? Wager too boldly and the player risks losing his entire uw The Mathematics of Gandling bankroll, even though he or she may have made only favorable bets, This #2 commonly known ae gamblers ruin. On the other hand, betting too conservatively the player severely limits his opportunity to make a good return on his capital. Fortunately for the player, there exists a mathematical theory for committing resources in favorable games. This will be discussed in Chapter 9. n Chapter 8 Mathematical Systems Before looking at the optimal strategy for exploiting a positive expectancy situation, it may be worthwhile o evaluate what I refer toas mathematical ystems. Although here I use rouleteas an example, the principles apply equally to craps and the Wheel oT Matematica syste” were he pay 5 system” Imean a stem wherethe pier decides which bet to male sng onl the flowing ifort ‘U)arecordof what nunbersbavecomeuponsomenumbe of pest spins, and (@) arecord ofthe bes he ha made if any. on hos sins. ‘We assime hee tat when te plas bet, fr him sl ambers are equally likly to come up on each spin ofthe whee. Tis sean not using bised wheels or pial prediction metho ‘Roulette har long ben the prototype of unbeatable embing sas. es normal regarded a a repeated independent is res which generate at each rl precisely one froma et Of Fandom numbers. Players may wager on particular subsets of Fandom numbers (ea te fist oven, even, an india rumber ct), win ithe number which comes upsantumber Of the chosen subset. player may wager on several subsets w The Mathematics of Gambling sirmltaneously and each bet is settled without references to the thers, TO fx the dscusiot let's consider the standard ‘American wheel. This hs thity-(18/38) x((40/38)" 40/38 —1) +$9004 (20/38) —$1 =S14.3645065. If we Give the expected loss by the average bet per cycle We Bet 8.756... =$14.36...1/19 exacly or ~ 5.26%. "These ealeulations ave tedious, and foreach system the details are different, $0 they have co be dove again. And there are an infinite number of gambling systems, so calculations can never ‘check them all out anyhow. Clearly this is not the way to under. ftand gambling ystems. The correct way ito develop a general 20 Mahenatel Sens ‘mathematical theory to cover gambling systems. That has been | done and here's how it works. First we define the wtion in & ‘specified et of betsto be the total ofall bes made, From what we Fave said, your expected (gun or loss is your action (Le, the tot ‘ofall our bets) times the house edge. For example, if you bet $10 pperhandat blackjack andplay for [Ohours,beting 1ODhands per hour, you havemadeathousand $1Obets, whichis $10,000worth ‘of “action.” If youarea poor blackjack payer and thecasno has| 123% edge over you, your exypeced losis $10,000 3% ~ $300, ‘Your actual loss may be somewhat more or somewhat les. I Nevada casino blackjack grosses a total of $400 ailon per year and the average casino edge over the player is 26 of the intial wage, then we can determine the total ation (A) per year: 2A =$400,000,000 so A =$20 billion. Thus from these figures ‘we would estimate $20 lion worth of bets are made per year at Nevada blackjack. The 2% Figure might be substantially off. We ‘could get a fasy accurate idea of che true figure by making a ‘careful statistical sumpling survey. I, instead, the igure is 4b, then A=510 billion. With 1%, A =$40 billion per year. Guidelines for Evaluating Systems “The general principles we ave disused apy to almost all gambling games, and when they apply, they gorante that Sms cannot ive heave an advaiase. "To help. you reject. ystems, here are condtions which auarante that a giem s worthless 1. Each individual bet inthe game has negative expectation. (Mis nike any sere of Bets have negative expectation ‘al iceman ito se ay posible pe is rules Out Systems Eke the no-imit doubling up sytem discussed) uns io TL The results of any one ply of the game donot “inftuence” the results of any ther play ofthe game. (Thus, in roulette, we assume that te chances are equally key forall ofthe numbers a The Mathematics of Gambling ‘on each and every future spin, regardless of the results of past spins) TV. There is a minimum allowed size for any bet. (This is necessary forthe techrcal steps inthe mathematical proof. Most people would take for aranted that thee is such a minimum, ramely some muliple of the smallest monetary unit, In the U.SA. theminirmum allowedbetissome multipleof one cent. In ‘West Germany, it may be some multiple of the pfenning, and so fot) ‘Under these condidons, iis a mathematical fact that every possible gambling system i worthless in the following ways (Any series of bets has negative expectation, (@) This expectation isthe (negative) sum of the expectations ofthe individual bets. {G) Ifthe player continues to bet, his total loss divided by his totalaction wlltend to get cover and closer tohisexpected loss divided by his total action. (It the player continues to bet itis almost certain that he wil: G@) be aloser; (b) eventually stay a loser forever, and so never again break even; (0 cventualy loses entre benkro, no matter how large ‘To give you an idea of how valuable this result is for spotting worthless systems, ere are some example of systems which can- ot possibly give the player an advantage: 1. All the roulette systems I have ever heard of, except the following wotypes. (a) Biased wheel, in which condition (1) may be violated; the numbers areno longer equally likely, so bets on some numbers may have positive expectation. (b) Physical prediction methods, in which the position and velocity of ball and Fotor are use to predict the outcome. 2. Allerapssystems {have ever heard of, except possibly those using ether crooked dice or physical ‘contra of dice, (Noe: While at the ith Annual Gambling Conference at Lake nm Mathematical Stes “Tahoe, I saw a dice cheat contol the dice, at a private showing {then baw him win ata easno. I heard he dit cegulaly. His badly mutilated body was found inthe Las Vegas area a year later, '3. Any systems for playing keno, slot and chuck-i-luck, ‘Asa further illustration, consider the book Garnbling Systems That WAN, pubiished by Gambling Tines, 1978, apes back, $2 Of the fourteen systems given there, our result applies at onceto eight. (The other six are one blackjack system, four raving systems, and a basketbal system.) (in the case of spors bets, itis generally dificult to determine ‘whether condition Iis satisfied. In the case of blackjack, condition [fais if the player counts cards and there ae, in fact, some win- ning systems, as most of you know) This eaves eight ystems in WAN: four raps systems, one bac- carat system, two roulette systems, and a keno system, CConeitions I through IV hold forall ight systems 0 none of them are winning systems. Nor do any of them reduce the house cedgein the sightet. However, they may be entertaining. Also, in ‘games tke keno craps, and roulette, where the expectation may vary from one game to another of from one type of bet tO another, some ways to bet are “smarter” (wransltion—less ddumb; more accurate translation—less negative expectation but sul losing) than others. Forthose whoare prepared tolose, but want tolose moresiowly, such systems may be of interes. ‘In mast cases, the basie information list of the various bets inthe game and their expectation. Then, if you mst play, choose only bets withthe least negative expectation, The "sjsem” com plexiies and hieroglyphic are not essential Temay amuse you fo see why coneltion TV is needed. Suppose, instead, tha there sno minimum bet and that we ae playing Red atroulette Our fist bet is $1,000, There is an 18/38 chance that ‘We win $1,000 and a 20/38 chance we lose $1,000. Now supose that the second bet is $0.90, de third bets $008, the fourth bet {is SOD, the fifth bet is $0000, ete. (Remember: no minima.) ‘Then the total ofall bes from the second on is $0.99999,..=$1.00 a The Mathematics of Ganbling ‘The total gun or loss on these bets i beween—$1.00 and + $1.0. “The toa action on all bets is $1,000 S1=91 001. Tf wewon the fist bet, our total winnings () will abwaysbe be- tween $999 and $1,001: This happens with probability 18/38, ‘Therefore, conclusions 4a), (6), andl (fal, Ako, ome tal action is $1,001 so T/A is always between $999/Si,001 and S1,001/$1,001. But our expected gain ()snewaive so E/Aisless than. Therefore if we win the first be, T/A doesnot en to get ‘over ane closer to E/A. Therefore, enclusion 3 aso fal ‘Conclusion 4(c) also deservessome comment. Actually thereis ‘an insignificantly small chance the player can win the casino’s bankroll before losing his. But even for moderate-size casino Dbankrollsthis possibilty iso tiny as tobe neglible, no mater hhow larg the player's bankroll! We will discus this in the next chapter. Is also discussed at some length in the 1962 edition of| iy book Beut she Dealer, and in Fller's great an Induction to Probability and its Applications, Vo. 1, Wiley. Thus, a more ‘exact version of conditions FIV would include information about the sizeof the casino bankrol. Then conclusion 4 would include information about the tiny chance that 4a), (), and (¢) don't hhappen. ‘Asfaras know, themost elementary mathematical proof ever ven for allthis is in my textbook, Bleranury Probab, available from Robert E. Krieger Publishing Co. Inc., 645 New ‘York Avenue Huntington, New Yerk 11743, The proof isoutlined on pp. 8485, exercises 5.12 and 5.13. It requires no calealus and ‘ean be followed by a good high school mathematics student i he ‘or she works through pp. 1-85. Wenow havea powerful tet for showing tata system doesn’t ‘win. Ths keeps us from wasing out money and ine buying oF playing losing systems. It als helps usin our search fr systems that do win, by greatly narrowing the possibilities. Chapter 9 Optimal Betting a_—_—SSS>=aaa_=—="! is somewhat rdicnlous to discuss an optimal money manage ‘ment strategy when the player has a negative expectancy. AS indicted in Chapter 8, wth an enforced house maximum and ‘minimum wager, thereisno way to convert negative expectation io a positive expectatlon through money manipulation. AMY ‘good money management plan says not to wager in such a situa tion. Players facing a negative expectancy should look elsewhere for a gambling game or, at the very least, bet insignificant mounts end wre ff ia their mind the expected fossa “entertainment.” ‘After the gambler has discovered a favorable wagering stua- Sion, les faced withthe problem of how best to apportion his limited financial resources. There exists arle of formula which you can use to decide how much to bet. I will explain the rule ‘and tell you the benefits that are likly if yo fllow it Le’ bepin wth a simple illustration that I deliberately exag- erated to better gt the idea acros. Suppose you havea very ich ‘adversary who wil let youbet any ammount on heads ateach ssf ps The Mathematics of Gambling ‘coin and that you both know thatthe chance of heads i some rhumber pgreater than Vs. your bet paysever money tie you hhave an edge, Now suppose p =0.52, so you tend to win 52 et~ cant of your bets and lose 48 percent. This is similar tothe tua tion in blackjack when te tn-count ratio i about 1.5 perent. ‘Suppose foo that your bankroll is only $200, How mich should ‘Youbet? Youcould play safe and just bet onecent each time. That ‘Way, you would have virtually no chance of ever losing your $200 and being put out of the game. But your expec gan s.04 por Unit or .Ofcents pe bet, At]00 one cet bets an hour, You expect to:win four cents per hour. I's hardly worth playing. Now lock at the o*her exreme where you bet your whole bankroll. Your expected gain $4 on the fist bet, more than if {you bet any lesser amount. f you win, ou now have $200. Ifyou ‘2eain bet all of ton Your second tum, your expected gain is $B ‘and is mote than i you bet any lose ‘susount, You make your expected gin the biggest on each tum by betting everything. But if ou lose once, you are broke and out ofthe ame. After many tums say 20, you have won 20 straight tosses with probability. "52" = 0000002090 and have a fortune of $104,857,600, oF you jhavelost once wth probability @.9999970/0and have nothing. In ‘general, asthe number of tosses increases, the probability that Sou will be nied tends to 1 oF certainty. This aes thestrategy of betting everthing unattractive. ‘Since the gambling probabilities and payoffs teach bet arethe same, it seems reasonable to expet thatthe “best” srateey will ‘alway’ involve betting the sae fraction of your bankroll at each tum, But what faction should this be? The “answer” isto bet fi — p)=0.52 ~ 0.48% 0.040 four percent of your bankroll cach tne, Thus you Be $4 the Dist tine, If you win, you have $104, so you bet 004 x S104 = $4.16 on the second tra. Ifyou lost the frst turn, you have $96, s0 you bet 0.04 x $96=$3.840n, the second turn Vou! continue fo bet four percent of your bbankrollateach tum, Thissratey of “investing” four percent of ‘your bankroll at each til and holding the remainder in cash is Known in investment circles asthe “optimal geometric growth 26 pina Being portfolio” or OGGP. In the 1962 eon of Bea the Dealer cusses aplication to scjack at som: kagth. Ther Tealled ithe Rely system, aller one othe malhematans who studied i, and I alo refered toi a5 (optima) fed fraction (of your bankzold betine ‘Why isthe Kelly system good? Fist, the chance of run s “sna” ta fat fmoney wete infinitely dvi ich tebe ite ue book eeping isead of actual ois and bil, o if we tc precious metas such gold or sv) them any stem whore You never bet everthing wil have zero chance of ruin because ‘ren i you lays lose, you sil have something le ater each tee. The Key sym has this Feature, OF cous, in sta prac te eois,bilkor chipsare generally sed andtheeisamaaimum sive bt Therefore, wih very unlucky Series of bts, one could ‘retualyhaveso lite ft hat hehastobetmoreofisbankeoll {ham the ste cals fr. For nts the mina bet were SH, then io our coin example, you must overbet once your ‘bansol is below 251 te minintn bet were onecent, hen YOU only Have to overbet once your bankroll below 25 ces If th bad luck then contin you cou be wiped ou. Te seoond dsiablepropery ofthe Kel ssemsthatfsome- cone witha sigan fren money management sytem bas Gn he same game, your otal bankroll wil prebably grow ater than is In fet, a thesameconinuesindefiaiely, yourbankzall eee by any reassigned multiple ne third sable property ofthe Kel systems that youtend toremhaspeied ev! of wings nthe leas averagetine For example, suppose you ar a winning eid counter a blackjack, fad you watt fo an your $40 bankroll up to $40,000. The brumer of hands you's have o play on average todo hs wil Usngthe Kall site, be very clase to the minimum possible us- ingany stem of money management. So summarize, the Kell system fs raively safe, you tend to have more profi, and you tend to gto Your goal inthe shortest tine. @ The Mathematics of Gambling Blackjack Money Management “The Kely system cals for no bet unless you have the advan- tage. Therefore it would tell you avoid games such ascrapsand Xenoandslotmachines, However, ifyouavethe knowledgeand shit gain an edgen blackjack, yuu ca usec Kel stem ‘optimize your tae of gain. The station in blackjack is more complex than the con oss game because (1) the payoff on aone- ‘nit nal et can vary widely, due to such thines as dealer ot player blackjacks, insurance, doubling down, pai splitting, ad Surrender, and (2) because the advantage or disadvantage to the player varies from hand © hand, However, we can apply di wi ss resus to blackjack by making some slight modifications. Fis, let's see where he coin tossexampe's bs. fixed fraction of four percent came fom. The ‘general mathematical formula forthe Kelly system this: In any {Ginel) fvorable gambling station orinvestment, bet that rac- tion of your bankroll which maximizes E In (! +f), where Eis the expected value and Ji the natural logarithm (co the base 71828...) Tis In fanxton is avalable on most band Calculators. In the case of coin tossing the bes fraction, which L call is given fora favorable bet by f*=2p ~ 1, where pisthe chance of succes on one tos, and 72, ie it the ames ether (at oo your disadvantage. Not oo that “'Tiscoineidenally your expected gain per unit bet. ‘Now your expected gain in blackjack varies fom hand to tha. Irwe think ofsueassivehandsas coin toses witha varying then weshould bet f*=2p — I whenever our card count shows thatthe deck is favorable, When the deck is unfavorable, we "should" bet zero, Ustor-type eam ola approximates thisideal ‘oftettng zero in unfavorable situations. Youcan also achieve this Sometimes by counting the deck end waiting unt the dock is favorable before placing your fist be. But itisimpractcal to bet zo in unfavorable situtot, so we betas sinall ass discreet. Think ofthese smaller, slightly unfavorable bes asa 'dain” oF ‘cax" which “water down” the overall advantage of the ne pial Beri favorable bets. To compensate for this reduced advantage, f* should generaiy be “sightly” smaller than the 2p ~ Zeomputed above. Another effet of the small, slightly unfavorable betsisto increase the chance of ruin a ite “The most important blackjack “correction” to the /* com ppted for coin essing is due othe greater variably of payoff. Peier Griffin caleulates thatthe “root mean square” payofT on ‘one-unit blackjack bet is about 1.13. It tums out then that /* should be corrected to about (2p ~ 1)/1-27or about 79timesthe advantage. Sade this 75 because of the “drain” ofthe small, ‘unfavorable bets and we have the farly accurate rule: For ‘favorablosituations at blackjack, itis (Kall optimal tobe a per cent of your bankroll equal to about 3/4 percent advantage. For instance, wth a $400 bankroll and a one percent advantage, bet 3/4 of one percent of $400, oF $3. ‘The Kelly System for Roulette In general in roulette, the house has the eda, and the Kelly sem say, “don't be.” Buc in my chapter on physical predic tion a oul, I described a method where we Shannon and), with theaid of an electronic device, had an edge of approximately ‘4 percent onthe most favored single number. That corresponds toawin probability of p= 0.04, with apayofT of 85 timesthe bet, and a probability of! ~ p = 0.96 of losing the bet. It turns out that += 44/95 ~.01257 The general formula for /* when you win A times a favorable bet with probability p and lse the bet with probability 1p, is f*=e/A where @=(A+ )p-1>0 the player’ expected gain Per unit bet or his advantage. Here Av35, p=.08, ad e= 0.4. In the coin toss example, A= J, P= 82, and e=.04. Using any fxed betting function f; the “growth rate” of your fortune is GO—p In d-=Afi+~pjin(t —P). Abr N bets ‘you illhave approximately exp(N_Gi/)timesasmuch money, “where exp isthe exponential function, also given on most pocket calulaters. The Mathematics of Ganbing For the roulette single number example, using my hand calculator (an HP6S) gives GO") = OOF In (1 + 357%) + 0.96 Im (1 ~ f°) =.04 In (1.48) + 0.95 In (0.98743) ~.08 x 36464 + 0.96 x (0.01265) =0.1459 ~ .01215=.00044, After 1,000 bets, you wil have approximately evp/2.44J= 17.47 times your starting bankroll ‘Notice the small value of /*. Thats because the very high risk cof oss on each bet makes it too dangerous to bet a arg faction ff your bankroll. To show the advantages of diversification, sup- pose insted that we vie our bet equally among the five most Favored numbers as Shannon and T2ctully did inthe casinos. fone of these univers come up, we win an amonnt equal to 5 774/Sof our amount bet, and if none come up, we lose our bet. ‘Thus A=31/5=6.2 The other four numbers are not quite aS favored asthebest number. However, toillstate diversification, ‘suppose thatthe Fivesway betas Une same 4 advange. TH ‘onesponds to p=0.20. Then f* =.44/6.2 = 0.07097, so youbet about seven percent of your bankroll and G")=0.20 In (+ 62/9) + 0.80 In (1 ~ J¥)=0.01d0d. This growth rate is about 5.75cimes that forthe single number. After 1,000bets, you would inave epprosimately 1.25 milion times your starting bankroll. Such isthe power of diversification, ‘What isthe price of deviating from betting the upd Kelly fraction ft tums ou that fr bet payofislike blackjack, which can be approximated by coin tossing, the “performance loss” is hot sors over several days play. But forthe roulette example, the performance loss from moderate deviations from the Kelly sgstem is considerable 0 APPENDICES APPENDIX A. eee eee Slaten een cteeantnas eed teietpeaece tee! ede Tureen tee eka ena Seg pos na Sal's Definition 1. Let F and G be probability distribution function. en lat spt an atom Peadeeney coop errbarees oe Pea Fenaeieced athe a et aay Conte te pine oer Seay if they hold for F and G. ae ™ Fe arinan Ou epalat oP 0 ae may dens cama ores eee nas de as Sire en Proper pee ote preen tame pcs fad gato ox ee Sorietnenentaametincanentan eae ee a The Mathematics of Gambling Homever probabilistic dominance is inadequate a a dafinition of “better because the typical situation is that Fis “spread out ‘more in both directions from the mean fall deck expectation E,=0. Thus F dominates G for x > F, and G dominates F for x < Ey, Infact G is (to 2 good approximation) a convex con- traction of F. More precisely, if E, and E, are the respective seans of F and G, we will find E, = B, >, with YE, a con- ‘ex contraction (thei equvalent othe notion “Tet is than” of pontolio theory); of X-- E,. Thus F is both “spread out smote” than G and translated in the positive direction more. The reason why E,, E 2 E, is because Ey i the expectation using the basic straiggy and constant bes, equivalent to the full pack expectation. When (advantageous) counting systems are used, the strategy for playing hands is improved whenever the player has ‘2en any cards other than tho ener he and the dealer use on the first round. Since this generally happens with postive probabil 'y, we then have E,, B > Ey ‘Definition 2. Point count system A is better shan system B if EE, and also POX =x) P(Y=) for x2. “Typically count systems satify E, > Ey = E, and X- (YE,), a = I (a special case" of convex contraction). ‘These conditions imply A is beter than B. ‘Assume that the betting ystems b(E) are numerical functions ofthe expectation B. Further assume b(E)=1 if Es 0 and b(E)= 1 if E>0. These are the ones generally considered. The popular fallacious systems such a the maringales (-. “doubling Up"), and the La Bouchere which incorporate past resulis, are of no inerest here. ‘Theorem 3 With the preceding notation and assumptions, if ‘A is bener shan B, then for any betting system by(E) based on the B point count, there is abetting system b,(E) based on the ‘A point count such thatthe return R, per nit bet by A (approx nately) probabilstically dominates R. Further, R, and Ry have approximately the same risk, Infact R,=R, te, Where €20. m Arpenices Proof: If F and ©. are continsous, define by by Bu (GUE)))=by I). Then oo tat he Ht unt ex ba fs expectation Ey for A and for B. The remainder ofthe bet non aro only i = Ey. Ten for coresponing percentiles ofthe respective distributions, A paces the sames bes a B. Bat FOE) = GUE) WE © Ey so A bas each isunce let a9 grat expectation, ence has at eat ax great expectation veal ‘Thos the ol expected retarn to Ai teat ts lare for B ‘Alo By = Ty per unit nace te bls paced ne the sane disibution Tn reality F and G ae ot continuous: insta they ae fie But they may be arbitarly closely spproximatd by contnooas distibtons so the result execs, with one qualition. If F {FG is daconnuoas, extend the gaps of Fand by adding Yer segments at the dicominsty points so that te extene Siono Fand @ have inves define on (0. Thea for tone Bsc that G is discontinooes at E” or Fis iscontinnous at FAG(E)) it may be necessary w define bE-(GIE))) “probably so ts maliple-alued, each valve oar ing with speci probebilis. “fo show that R=Re, which implies he same sk it sue fees wo assume that each prcetle level y forthe dita thom Fand G we hve the conitionl dso given sts ‘ying Fly) Got) where 9) = 0 Since tis onl olds approximately in practice, we have Ra=Ry Now we tm othe problem of mestring “closeness of @ given coin tothe “ultimate egy We sl ase tht point unt strategies ae ofthe form CCG) whee i the Yale assigned fr an 28, egy fe the Pine cOUNS FOF Tinks 2 trough 9, and eg=* me) ae he point coun for tems, cs, queens and Kings eopoctvely.Inpratse these are Tunpod togetir and only ten point count vale are specie. By writing C wih 13 components we aun asymmetry which yds substantially simpler proofs. Noe that C and 3, ae eqivaleat and wil be ented ‘The Mathemates of Combing Definition 4. EA E.=0 the ulimate strategy U=(0,.---0) is the one given by =A F, where AE, i the change i expec: tation from removing one ith card from the complete pack. If 'A E,=0 then U is given by upd. In Table 2.2, we have d for ane deck i¢ 024 and for four decks is O17. The u, rows are calculated in Table 2-2 from Definition 4 Tis tempting 1o think of U as representing to good approxima- tion the aiection of the gradient Fat f= ---=f,=V/3 of the player's expectation EXC...) as a function ofthe fraction f, of the cards from i=! to B. Then we calculate (C) TCHSLICHSIUL. i. the projection of Cin the E direction ‘The numerator isthe ianer or scalar produet and ICI =(Ee2)*. ‘Next we claim that d (C) gives the approximate ratio of the spread ofthe C distribution F, about Et the U distribucon F, bout E,, Then d (C) s the Genie tasure of closeness. I particula, for approximately the same risk per unit, and the same ‘istrbution of the bet sizes, it would follow that E(R) = F(R)i(C). Then C, and C, are arbitrary strategies EiR, VER = N(C)/A (C) for the same risk level and distribution of bet siaes. Thu the “power” ofa strategy is proportional wits (C).. ‘This conclusion is true but the argument must resolve 190 ‘obseaces: () In the preceding discussion we treated C, U, VE, ete, as though they were given in Cartesian coordinetes when in fact they {@) The probability distribution of E(,...f, must be con- sidered in reaching the conclusion and in general wil ivaidate it "Note further that both U and C are linear approximation to ‘an in general curved surfice”. Also in the real case ue domain ‘is. Lage finite subset of poins of the posible (ff). each ‘of positive probability. (The original discovery of winning back- Jack sytams (Thorp, 1961 was motivated by tis model.) Fist introduced the Ein...) "surfoce”, where o, i the number of cards remaining of denomination i, Intuitive arguments “con- Mm Arpenices son meats hie ssh ora oo, mae met ne ict aes “Ty ms ho ee edict wins dn See eit NTL he emt oh See pletion Telerik meee ee ai ay a a an fait yarn tea i ieeeen eet (ise ctutrmc lt aetna Nene eee armas HELA See EELS ah pata eae The Mathemais of Gambling APPENDIX B. ‘Suppose (Hypothesis 1 thatthe shoe really has four cnuphete decks, Then the number X of unseen tenoalue cards among the 14cards (wo decks) not seen will average 32. Inthe general case with L unseen cards, T tens in the whole pack. and N non fn in the whole pac, the average value A of Xisgiven by A= UT/(N + 7). In our example, U=104, T=64 and N~= 14, so we ect A= 104 64/208 ~32. But there wil be a fluctuation around this number. Mamnemadcians use Ue standard deviation $ to ‘measure this fluctuation. The formula S*=/UTN/T + N)'J0 ~ (U~ WiN+T~ Uh ‘Foro example, S°=/104 x 64 x 144/2087)(1 ~ 103/207) 1304, 30 5 = 11-B0¥ = 3.3362. To a good approxima- tion, X is “normally distributed” with mean A= 32 and stan- dard deviation $ = 3.3362. Now, Suppose nse (Hyptless thatthe deck hasten ten~ value cards removed. Then U=9, T= Sand N=134.If Yisthe ‘number of unseen cards, we have the real A'=25.6364, but we think there are ten more ten-value cards. So assuming incorrectly that no ten values are gone, the number tha we deduoe for Y has aan average of 4 + 10= 35,6364, The real for ¥is 9 x 54 x BAI9B (1 ~ 931897) = 91593, so $= 30264. ‘What e wal o kiwow fs whether t0 Bele Hypothesis T (null hypothesis") or Hypothesis I. This isa clase statistics, problem, Itturnsou that inorder forusto havea good chanceto believe the corecthypothess the Avaluefor X and Yneed tobe at least two and preferably several Sunits apart. Inthis example, they differ by only 35.6364 ~ 323.6364 which is about one $ ‘unit, Ofcourse, repeated countowns ofthis same shoe wil again {increase Gur ubliy 1 tel wheter the shoe is short. 18 APPENDIX €. or this nt spe dain, es supp atu expC) +, where a,b, and e are constants and exp i the exponential funcon. This on ofthe simplest material fncons that haste right shape" (No Fedo tis dscuson ws te uate 3 ‘Ste dierae) that th ball velocity atthe point where el fom track ws about 0. revlons per cond ps) aid ttt revolutions exert was about 2p. Use is andthe choise {0 when the ball leave the tack gies a= 10/3, b= 3720 and 5 13-3) 1nd ts ivesan angular cacy vintps fv) = hep, Faure HT shows a graph of x) eae APPENDIX D. A calculation shows, fr our istrative (function that aff)=Mewist 20) ~ I) 10/3. Tos from t we can pred ‘Hemant freon ul he bal aves the ack Fore Stance, it 7 / ec, we pratt the bal wl lee the tak in Gia eg(ad) I~ ford= 285 eats aie he ‘irchichitthesecond ie. ned 7~ see then we red fA) =9.51 revolutions. cee The Mathematics of Ganbling APPENDIX E. Math readers: de(DdT= - (3x) +103/60. It can be shown that forthe «of this example, the error A xP in the prediction of (7) due wo an error & T in measuring 7s given by 8 sf) =~ Ons) + 10PTCO = —3 T/OlespCT20) 7p), For instance, if T = 08 sec. and AT =0012 Sec, we have a prodicton error of x,(08) = Gl revs oF 4.2 numbers ‘on the wheel In our illustration T'= (8 sec. means x,(7) = 451 revolutions to go. The time to go is 20/)log.33,00 + 1) or 570 sec. We have somewhat less time than this to bet APPENDIX F. In our example, the equation for 1,7) is_ 17) OOP jog 30) exp(2T720) = I) } = (20/3/68, 2x,(TVA0 + 1). The error is approximately £7) = ~(ATexp(37?20) Hesp(3020) ~ 1). Thus again if T= 08 sec. and A T= 0012 see. A(T) = ~@106 sec. With a rotor speed of 0.33 rps {his causs a rotor prediction error of WO rev. oF 1.3 pockets. Inour example then, we measured Too larg by 0012 see. This led us to believe the ball would leave the wack ata point about 442 pockets before whore it did. Therefore, we forecast impact 6 the rotor 42 pockets early. It also led us to believe the ball ‘would lave the track sooner in time, Thus, we thought the rotor ‘woulda revolve as far a8 it did. This made us forecast impact nodner 13 pockets early, fora total error of 5.5 pockets early ‘There are other important sources of errr, so out final predic- tions were not this good. But they were good enough, Tn summary, ote that an error where AT is positive, ie, we think T's bigger than i really is because we ht the switch early the frst time or late the second time, leads us to think the ball a8 Appences fs slower than its, That makes us think 27) is shorter. Thus, \we expect the ball atthe rotor to soon and forecast impact on fe rotor ahead of where it tds to eccut. Comversely i Tis ‘negative (last onthe first witch or early onthe second), we think Tis smaller, the ball is faster, and mistakenly forecast (7) and 4,(7) a8 too big, Then we predict impact behind where it tends 10 occut, ‘The rotor angular velocity, followed a law close 0 ft) = Aexp (0). A typical value for A was 0.33 rex./sec. The “decay” oF ‘dlowing down" constant b was vory smal. The rotor is massive and spins on a well-oiled bearing (on our casino whee, it was the ‘pointed end ofa sturdy see shaft). Inthe course of a minute oF two, the slowing was hardly perceptible, (Note: Siroboscopic “bea frequency” techniques, plus an accurate clock, can quickly ‘and easly give avery revise measurement of b and te lowing, down.) Let's take b=—Log,101)120 or 0000704oee., which cor ‘responds toa slowing down from 033 rev sec. t9 0.30 rev./se. in ‘twominutes. This seems like the right order of magnitude. To put the rotor position into the tiny computer we were going to build, ‘we planned toh arotor timing switch once when the zero passed. reference mark on the wheel, and then hte switch again when the zero passe the reference mark a second time. Since the rotor ‘velocity war small and nearly constant, this was ales seneitve™ imeguremen. Therefore weplanedio doit is shorty bore the ball was spun, How much error in the ball’s final postion (pocket) comes, from rotor ming errors? Assume for Ssmplicty that the rotor ‘makes one revolution in about thre seconds (33 rev./sec.) and that we can neglect the sowing down of the roto. Then, asin the ball timing, we might expect atypical (root mesn square) sie of| about 11.2/1,000 seconds forthe combined effet of the two er rors. If te rotor really makes one reolution in 3.000 seconds, and we think it takes 30112 seconds, then in 30 seconds we think the whee! wal travel 49628 revolutions whereas t realy travels 10,0000 revolutions. Thus, the rotor goes 0372rev. oF LA pockets farther than expected Simlaly, if e think therotor takes 2.9888 a9 The Mathematics of Gambling seconds for one revolution, then in 30 seconds the rotor goes 10375 1ey. oF 1 pockesess than we expected. APPENDIX G. Tam using the normal approximation forthe statistical discus sion. I think itis very nearly an accurate description of what happens and that this approximation only slightly affects the discussion APPENDIX H. In genera, there are exactly (5+2!/5e! home board positions with exactly r men, There are exactly (6+)!/6!e-1 home board positions with from one tor men, Thus, since r=15 is posible fhe awa yan, hove ave a Wal of 216! 18! 1=54,263 di ferent home board postions for one player. The symbole, read “tfacorial.” means LX2x3X. xf. Thus 1, 2!=2, 3-6, 41 Scholarly References For thace readers who are expecially interested in the technical \work behind the material in this book and other work by Pro- fessor Thorp, here is a list of some of his related scholarly publications. Game Theory 1. “A Fivorable Strategy for Twenty-One," Proceedings ofthe National Academy of Scences 47 (961). 0-2. (MR #B846). 2. BOOK: Beat the Dealer: A Winning Strategy forthe Game of Twenry-One, Random House, New York, 1962; Vintage paper- back also, 1966; revised edition, 1966. 3. “A Partial Analysis Of Go;" (with W. E. Walden), The (British) Computer Journal 73 (1968), 203-207. (MR33 #2424), @ 123, 340). 4. “A Favorable Side Bet in Nevada Baccara,” (with W. B. Walden), Journal of the American Statistical Assocation 6, Part 1 9966), 37-328 5. “Repeated Independent Tilsen a Class of Dice Problems,” (Mathematical Now), American Mathematical Monthly (August September 1964), 778°781. Z 123, 364). 6. “Optimal Gambling Systems for Favorable Games," Review of International Ins of Stiscs 31:3 (969), 7-293. (2.18, 4m, a 7. “Solution of a Poker Variant.” Information Sciences 2/2 (0970), 299-301. (MR S42 48952), (Z'205, 232). ‘8. “A Computer Assisted Study of Go on M x N Boards” (with W. E. Walden), Lecture Notes in Operations Research and Mathematical Systems in Theoretical Approuches 1 Non- ‘Numerical Problem Solving, Vol. 28, R. Banerji and M. D. Mesarovc, eds, Springer eriag, New York, 1970, 303-343 later revised version, Information Sciences 41 Q972), 133 (MR4S 8684), (2 22868028). 9, "Non-Random Shufing With Applications tothe Game of Faro” Journal ofthe American Statistical Association, 842-87, ‘Deceniber 1973. Much expanded version appears in Gambling ‘and Society, edited by W. Eadington, Charles C. ‘Thomas, Springfield, Hinois, 1975, as: Probabilities and Strategies forthe ‘Game of Faro, pp. $3160. 10, “The Fundamental Theorem of Card Counting With Ap- plications to Trent et Quarante and Baccarat." (with W. E- Walden) Intemational Journal of Game Theory 2 (973), 108-19. (@ 25890058) UU, “Backgammon: Part 1, The Optimal Strategy for the Pure Running Game." Proceedings of the Second Annual Conference ‘on Gambling, Lake Tahoe, 17S. mis. 42 + pp. See news report, fon this paper in: “Beating the Game,” an article by Dietrick E. ‘Thomsen, Science News, Vol. 107, March, 1975. 12, "Blackjack Syetems:” Proceedings of the Second Anal CConferrence on Gambling, Lake Tahoe, IS. ms. 5 + pp. See sews report on ths paper in: “Beating the Game.” an article by Dietrick E. Thomsen, Science News, Vol. 107, March, 1975. Probability, Statistics 13. BOOK: Elementary Probably, Wiley, New York, 1966. Mathematical Finance 14, BOOK: Bea the Marker, (with S. Kassou), Random House, New York, 1967 a 1S, “Portfolio Choice and the Kelly Criterion” Proceedings of the 1971 Business and Economics Section of the American Staitical Association 1972, 215-224 To be reprinted in Invest- ‘ment Decision-Making, edited by J. Bicksler. Reprinted in Stochastic Opdnisation Models in Finance, Acadetnic Press, edited by W. T. Ziemba, S.L. Brumelle, and R. G. Vickson, 191, pp. 599-620. 16. “The Capital Growth Model: An Empirical Investigation: (with James Bicksler), Journal of Fnancial and Quanivatve Analysis, March 1973, Vol. VI, No. 2, pp. 273-287. General Interest 17. *A Professor Beas the Gamblers", The Ailanic Monthly, June, 1962. Reprinted in “The Gamblers Bedside Book,” John K. Hutchens, ed., Taplnige, New York, 1977, pp. 1665177.

You might also like