Duarte 1 Gavin Duarte Math 20 Hamaker 30 November 2011 Blackjack stirs interest from all types of people, from the blackjack pioneering mathematician Edward Thorpe to the vacationer traveling to Las Vegas for the weekend.1 Blackjack intrigues people because it is the only gambling game where the player actively determines his or her odds, and one’s knowledge of strategy can greatly affect the probability of winning. Because of the ability to alter the probabilities based on skill level, it is possible for an experienced player to have a positive expected value of winning. In order to fully comprehend the strategies behind blackjack, one must know the rules of the game and how they affect play; in fact, most of the player’s options can be used to the player’s advantage. The overall premise of the game is for the player to beat the dealer by obtaining a sum of cards, a hand, less than or equal to 21 and at the same time have a higher hand than the dealer. The game consists of one dealer and up to seven other players, and the number of card decks, denoted by n, which ranges from n equals 1 to 6 held in a shoe. All number cards equal their respective values, and each face card (jack, queen, king) is valued at ten points. An ace can be worth one point, denoted a hard ace, and eleven, a soft ace, whichever value makes the player’s hand better. Blackjack occurs when the first two cards dealt to a player are an ace and a card valued at ten. Blackjacks are paid at 3/2 the original bet. The game starts with the dealer dealing two cards to each player from left to right, the dealer last, and in turn, 1 Thorpe wrote Beat the Dealer: A Winning Strategy for the Game of Twenty-One (1962), the first attempt to mathematically prove correct strategies for blackjack. Duarte 2 each player can stay (keep the hand they have), or hit (draw another card) as many times as desired. A player busts if the total is over 21, and automatically loses the hand to the dealer. The rules for this game will be similar to those used in Las Vegas, having the following criteria. A player initially bets a number B for certain hand, and may double down (place double the original bet) upon receiving the first two cards. If the player receives a pair, they may split the hand into two separate hands and place an additional bet, B. If the dealer’s up card (every hand the dealer’s first card is placed face up) is an ace, the player can make an insurance side bet of π΅⁄2 that the dealer gets a blackjack, and if this happens, the player gets B returned to him. The rules for the dealer are very specific; they must continue to hit until they have a total of 17 or greater, whether or not the dealer busts. Why does the dealer have the edge? The reason the dealer has the advantage is that they are always the last person to play, so everyone at the table has a chance to bust (and therefore lose the hand) before the dealer has to play. The Basic Strategy No matter what minor rules variations are present in each specific game of blackjack, it is important that one remembers the basic strategies that help improve the player’s odds (Uston, 47).2 These points are often determined by the dealer’s upcard because that is the only piece of information the player knows about the dealer, and the goal of blackjack is to beat the dealer. 1. If the dealer’s upcard is a 2 or 3, hit until you (the player) at least a total of 13. For example, if the dealer had a 2 upcard, and your hand totaled 12, you would want to hit. 2. If the dealer’s upcard is a 4, 5, or 6, hit until you have at least a total of 12. The reason the number total for the player is less than in part 1 is because the dealer has a 2 These strategies may be slightly different if one is using advanced counting techniques instead of a basic strategy. Duarte 3 higher chance of busting given these upcards, because the dealer must hit until 17 is reached. For example, let ππ denote the value of the π π‘β card for the dealer. So π1 = 6 means the dealer’s upcard is a 6. 10 π(ππππππ ππ’π π‘π | π1 = 6) = ∑ π(π2 = π) ∗ π(π3 > 15 − π | π2 = π) π=6 5 + ∑ π(π2 = 5) ∗ π(π3 = π) ∗ π (π4 > 4 + π | π2 = 5, π3 = π) π=1 5 + ∑ π(π2 = 4) ∗ π(π3 = 1 + π) ∗ π (π4 > 4 + π | π2 = 4, π3 = 1 + π) π=1 + π(ππππππ ππ’π π‘π | π2 = 3) + π(ππππππ ππ’π π‘π | π2 = 2) + π(ππππππ ππ’π π‘π | π2 = 1) = .42 So the dealer has a 42% chance of busting if their upcard is a 6, and this is why the player needs only to total at least 12, because the probability of the dealer busting is so high. 3. If the dealer’s upcard is anything 7 or higher (7, 8, 9, 10, or ace), hit until you have at least 17. The reason is that the dealer has a higher probability of having a good hand and not busting with this upcard, and if you have any hand lower than 17, the only way you will win is if the dealer busts. Finding the probability of these situations is similar to the way we found the probability in Part 2, where the dealer’s upcard is a 6. 4. Double down if you have a total of 11 and the dealer’s upcard is not an ace (i.e. 210). This is because the probability of you having a high hand is greater because you Duarte 4 only need one ten valued card to reach 21, and even a card such as an 8 or 9 will still give you a good hand. 5. Double down if you have a total of 10 and the dealer’s upcard is any card from 2 through 9. If the dealer and you both get a 10 valued card, you will still win the hand, and if the dealer gets an ace, you will push (bet goes to the next round). That is why you double down with a 10 or 11 only when your total is at least one number greater than the dealer. 6. Split pairs when you have 8s or aces. The thought process behind this is that if you split aces, you are turning a hand worth 12 into two very good hands both starting at 11. If you have two 8s, 16 is a bad hand because you will only beat the dealer if he busts, and instead you give yourself a better chance by starting off with two separate hands. 7. Never take insurance. If you have a blackjack, you might think of taking insurance so that you will always win at least your original bet back. But this only occurs if the dealer has a blackjack (natural) as well. You are better off taking the probability that the dealer will not get blackjack, and then you will win 3/2 your original bet, because the probability that the dealer has blackjack given that you have blackjack is smaller than the probability that only you get blackjack (and therefore win 3/2 your original bet). The Advantage of Card Counting Many astute mathematicians have developed complex ways of card counting in order to benefit the player’s odds. Card counting consists of making a specific value (point-count value) for a certain type of card in each deck. This works because each card is a dependent trial, therefore knowing more information about the previous trials will lead to a better understanding of the current trial. First, we must define EOR, or the effect on the player’s expectation of Duarte 5 removing individual cards. That is to say, when a specific card value is removed from the deck (already played), the player’s expected value of winning changes. If the number is positive for the respective card, it is advantageous to the player when this card is removed. Table 1 below gives each EOR value by percentage change.3 Card Value EOR 2 3 4 5 6 7 8 9 10 Ace 0.381 0.434 0.568 0.727 0.412 0.282 -0.003 -0.173 -0.512 -0.579 Looking at these values, one can clearly see when cards 2-6 are removed, the players expectation is significantly higher than the mean (µ=0), whereas if a ten or ace is removed, the expectation is significantly lower than the mean. This can be interpreted in a more common sense approach. If more low cards have been removed from the deck, and the dealer has a hand of 12-16, there is now a greater probability that the dealer will draw a high card and bust. In terms of different card counting techniques, one must determine the playing efficiency and betting correlation of each technique, as well the overall ease of use. Betting correlation can be computed by taking the dot product of the point count values and the EOR values, and then dividing by the square root of the variance of the point count values and EOR values. This is essentially the same as the standardized sum, because you are taking the values compared to mean and dividing by the standard deviation. So, the betting correlation is π΅ππ‘π‘πππ πΆπππππππ‘πππ = π· , where Φ is the √(π£πππππππ ππ πππππ‘ πππ’ππ‘ π£πππ’ππ )∗(π£πππππππ ππ πΈππ π£πππ’ππ ) dot product of the point count values and EOR values (an example will be given later).4 The playing efficiency is computed by the quotient between the expected gain from the lack of cards remaining in the deck and the hypothetical gain if the player had knowledge of all Values obtained from Schlesinger’s Blackjack Attack Note: The Theory of Gambling and Statistical Logic also gives equations for finding the SCORE, which combines the values of betting correlation and playing efficiency. 3 4 Duarte 6 the cards. This is very similar to the cumulative distribution function that we learned in class, except that it is the bivariate normal distribution. So, the expected gain, EG, without using a basic strategy is πΈπΊ(π) = ππ 2 ⁄2π 2 π √2π± π −µ − µ √2π± ∞ ∫ π −π₯ 2 ⁄2 ππ₯ µ⁄ππ where µ is the expected gain using only basic strategy (see section The Basic Strategy), and ππ is the standard deviation for the mean of an n-card subset of the deck, given by ππ = π2 51−π √ π ( 50 ), which essentially accounts for the change in the standard deviation after a certain number of cards, n, have been played. The hypothetical expected gain from using a point count system is also bivariate normally distributed with correlation coefficient ρ.5 Therefore, the hypothetical expected gain, π»πΊ(π) = πππ √2π± π −µ2π ⁄2π 2 ππ2 − µπ √2π± ∞ ∫ π −π₯ 2 ⁄2 ππ₯ µπ ⁄πππ where µπ and ππ are the mean and standard deviation values using the point count system. So, the playing efficiency is simply π»πΊ(π) πΈπΊ(π) The higher the betting correlation value, the lower the playing efficiency, because they are inversely related, so if a certain card counting system has a high betting correlation, it will most likely have a lower playing efficiency. Hi-Lo Card Counting Strategy 5 Bivariate distribution functions have correlation coefficients defined under parameters defined at http://mathworld.wolfram.com/BivariateNormalDistribution.html Duarte 7 One card counting technique that is fairly easy to understand and implement is the Hi-Lo system. This assigns a value of +1 for cards valued 2-6, 0 for cards 7-9, and -1 for a ten or ace. Starting at zero, when a card is card is dealt the player simply keeps a running tally in their mind about the total number by adding 1, 0, or -1 to the total. This system is considered to be balanced since the total equals zero after all the cards are dealt because the expected value, πΈ(π₯) = ππ, π π πΈ(π₯) = −1 ∗ π(ππππ 2 − 6) + 0 ∗ π(ππππ 7 − 9) + 1 ∗ π(ππππ π‘ππ, πππ) 20 12 20 πΈ(π) = −1 (52) + 0 (52) + 1 (52) = 0 Table 2 below illustrates the values for the cards. Card Value Hi-Lo Point Value 2 3 4 5 6 7 8 9 10 Ace +1 +1 +1 +1 +1 0 0 0 +1 +1 The betting correlation of the high low system would be the dot product of Table 1 and Table 2 divided by the product of the variances of table 1 and table 2. So, π΅ππ‘π‘πππ πΆπππππππ‘πππ = πππππ 1 β πππππ 2 √(π 2 ππ πππππ 1) ∗ (π 2 ππ πππππ 2) = 5.15 √10 ∗ 2.842 = 0.966 because µ for both Table 1 and 2 equals zero. The way the player uses this technique to his or her advantage is that if the total is greater than 0, the player is at a slight advantage at that point in the deck, and the dealer has better odds when the total is less than 0. The reasoning behind this is that if the Hi-Lo value is much greater than 0 (e.g. 6), there is a much greater probability of being dealt a ten or ace, which allows the player to have a stronger hand. Also, it makes the dealer bust more often because they must hit on hands valued 12-16, which would therefore bust with a higher probability. Duarte 8 Some downfalls of card counting are that a player will not be able to see every card being dealt, and therefore the Hi-Lo value will never be exact as the decks diminishes. Also, the advantage to the player can be lessened if a casino uses multiple decks of cards, uses burn cards (cards placed face down before each hand), and reshuffling. The different tactics used by casinos cause the player to have a disadvantage. Consequently, the fewer decks being use (1 is ideal), and the more cards the player is able to see being played, the more effective card counting becomes. Other Ways of Modeling Blackjack- Markov Chains Because blackjack uses cards without replacement, each subsequent card is dependent on the cards played previously from the deck. This can make computations very difficult, and that is why many of the values obtained for the game are approximated by doing a tedious amount of simulations. However, if one thinks about blackjack in terms of moving from one value to the next with a certain probability, from one state to another, Markov Chains can be very efficient at modeling the complexities of blackjack.6 Because the dealer has specific rules to follow, a hand, H, which is 17 or greater, can be thought of as an absorbing state, because the dealer must stand on those hands or have busted. The state space representing the dealer consists of elements where π» = 2 − 11, where the first card is valued between 2 and 11 π» = 4 − 16, where the first two cards total to a value between 4 and 16 π» = 17 − 20, where the dealer must stand π» = 21, where the dealer has blackjack (only on the first two cards drawn π» > 21, where the dealer busts 6 A paper written by Rice University engineers called A Markov Chain Analysis of Blackjack Strategy Duarte 9 (Note- Assume the dealer stands on soft 17) Because the dealer must stand if π» > 16, the last 3 elements can be described as absorbing states, because the probability of reaching that state again is 1. If one makes a transition matrix, P, the dealer will reach an absorbing state in at most 17 steps (i.e. π17 ), because the value must keep increasing until it reaches an absorbing state. The same idea can be applied to the player’s hand, with the additional options of splitting pairs, doubling down, and the ability to hit on any total less than 21. This leads to the player reaching an absorbing state by at least π21 , because the player has more options (states) from which to choose. The player’s expected return on a given hand can be determined by taking the average of the player’s profit with both the player’s and dealer’s absorbing states given initial card values. Markov chains can also be used for card counting systems. For example, the Complete Point-Count System uses an ordered triple, (Low, Medium, High), which has the same criteria as Table 2. Using the ordered triple, the number of cards remaining in the shoe, R, can be represented by π = 52 ∗ ππ’ππππ ππ πππππ − πΏ − π − π». This, in turn, leads to a formula known as the High-Low Index (HLI) π»πΏπΌ = πΏ−π» ∗ 100 π This formula is much more useful than the Hi-Lo Technique for changing your betting amount. Edward Thorp gives the following cases for betting given a uniform starting bet B and a given HLI range: [−100,2] π»πΏπΌ = {(2,10] (10,100] πππ‘ π΅ π»πΏπΌ πππ‘ π΅ 2 πππ‘ 5π΅ Duarte 10 Common sense tells us that a high HLI favors the player (because there are more high cards in the deck, the same as the Hi-Lo system) and therefore a player should bet more under those conditions. This provides a slightly greater expectation for the player because they can alter the betting amount as more cards in the deck become known. For example, if the dealer has dealt 8 low cards, 4 medium cards, and 2 high cards in a one deck game, the π»πΏπΌ = 8−2 ∗ 100 = 15.8 38 so the player should place a bet of 5B for the next bet. This leads to finding a distribution for a single hand, D, which is the total of all possible distributions of each card type. So given a card value, for example a 2, would have the distribution π(2) = 1 20π − πΏ ∗ 5 π where N= number of decks. We multiply by 1/5 because 2 is one of 5 low valued cards, and there are 20 low valued cards per deck. The same approach can be applied to both the medium and high valued cards. Using the distributions for the specific card values, one can average the triples and create the probability of the player’s advantage using Markov chains by starting at a specific state in the transition matrix (depending on the dealer’s upcard and player’s hand) and then determining the probabilities of each specific outcome (bust, win, etc.). Conclusion A small change in the rules of blackjack can have a profound effect on the expected winnings for the player. However, learning the basic strategy behind blackjack allows anyone to better themselves without much prior knowledge about math. However, the dealer will continue to have an advantage unless the player uses more advanced techniques. If the player uses card counting techniques, the expected profit can reach +0.02 if using only one deck, where 0 is Duarte 11 breaking even.7 However, the more decks that are in the shoe, the less the player has an advantage. A player with no real experience has an expected loss of -0.08, while the basic strategy will get you to around -0.01 to -0.03. Clearly, the percentage changes are small, and that is why these techniques must be implemented for long periods of time in order for them to be useful. Even when using the more exact Markov Chains to find the probabilities, it still only yields an expected gain of +0.02. Even though some basic assumptions are made to simplify the computations, the strategies still proved to be very accurate to the actual probabilities. Because each successive hand is dependent on the previous hand, finding the exact probabilities can get rather messy (as seen in Basic Strategy, Part 2). This is why many of the numbers are found using simulation over hundreds of thousands of times. Using the player options such as splitting and doubling down is the only way for the player to create an advantage, and the player must rely on their own intuition to make these decisions. However, blackjack remains the only casino game where the player can have the advantage, and models such as Markov Chains help to illustrate the math behind the game. 7 Formulated in Million Dollar Blackjack via numerous simulations Duarte 12 Appendix The probability of the dealer busting given the dealer’s upcard value.8 This correlates with the section Basic Strategy, Part 2 Dealer’s 2 3 4 5 6 7 8 9 10 Ace Upcard P(Dealer 0.353 0.3756 0.4028 0.4289 0.4208 0.2599 0.2386 0.2334 0.2143 0.1165 Busts) The probability of busting while taking a hit. This is fairly straightforward. For example, if you have a hand equaling 20, the only way of not busting is if you are dealt an ace, so π(ππ’π π‘ | π‘ππ‘ππ = 20) = 1 − π(ππ ππ’π π‘ | π‘ππ‘ππ = 20) = 1 − 4/52 = 0.92 Hand Total P(Busting) <11 12 13 14 15 16 17 18 19 20 21 0 0.31 0.39 0.56 0.58 0.62 0.69 0.77 0.85 0.92 1 The overall expected profit of the player if using the Complete Point Count System (in the Markov Chain section) in terms of increasing deck sizes9 Number of Decks E(Profit) 8 9 1 0.0296 2 0.0129 4 0.003 Table from http://www.blackjacktactics.com/blackjack/odds Table gathered from results of Markov Chain Analysis in A Markov Chain Analysis of Blackjack Strategy Duarte 13 Bibliography "Blackjack Probability Odds - Winning Blackjack Odds Charts." Play Casino Blackjack Online Basic Strategy and Free Blackjack. Web. 29 Nov. 2011. <http://www.lolblackjack.com/blackjack/probability-odds/>. "Blackjack Odds." Blackjack Strategy. Web. 29 Nov. 2011. <http://www.blackjacktactics.com/blackjack/odds/>. Epstein, Richard A. The Theory of Gambling and Statistical Logic. Burlington, MA: Academic, 2009. Print. Griffin, P. A. The Theory of Blackjack. Las Vegas, NV: GBC Pr., 1979. Print. "Hi Lo Card Counting System." Blackjack Strategy. Web. 29 Nov. 2011. <http://www.blackjacktactics.com/blackjack/strategy/card-counting/hi-lo/>. Schlesinger, Don. Blackjack Attack: Playing the Pros' Way. Las Vegas: RGE Pub., 2004. Print. Thorp, Edward O. Beat the Dealer: a Winning Strategy for the Game of Twenty-one. New York: Vintage, 1966. Print. Uston, Ken. Million Dollar Blackjack. Secaucus, NJ: Gambling Times, 2003. Print. Wakin, Michael, and Christopher Rozell. "A Markov Chain Analysis of Blackjack Strategy." Thesis. Rice University. Print.