Presentation on theme: "Mathematical Finance Seminar"— Presentation transcript:
1Mathematical Finance Seminar The Kelly criterion and its variants: theory and practice in sports, lottery, futures & options trading The symmetric downside Sharpe ratio and the evaluation of great investors & speculators and their use of the Kelly criterionWilliam T ZiembaAlumni Professor at Financial Modeling and Stochastic Optimization, Emeritus, Sauder School of Business, UBC, Vancouver, BC, Canada V6T 1Z2Mathematical Finance SeminarUniversity of ChicagoApril 6, 2007
2ReferencesMacLean, L C and W T Ziemba (2006) The Kelly criterion: theory and practiceThorp, E. O. (2006). The Kelly criterion in blackjack, sports betting and the stockmarket.in S A Zenios and W T Ziemba, eds.,Handbook of Asset and Liability Management, Volume A: Theory and Methodology, North Holland.MacLean, Sanegre, Zhao, Ziemba (2004) Capital growth with security, J.Economic Dynamics and Control, How to calculate the optimal Kelly fraction subject to being above a wealth path with high probabilityZiemba, W T (2005) The Symmetric downside Sharpe ratio and the evaluation of great investors and speculators, Journal of Portfolio Management (Fall).Chapter 6 of W T Ziemba (2003) The Stochastic Programming Approach to Asset Liability and Wealth Management, AIMR updated into various chapters in Ziemba and Ziemba (2007), Scenarios for Risk Management and Global Investment Strategies,Wiley, July which is Wilmott columns merged into a bookSamuelson, P A and W T Ziemba (2007) Understanding the finite properties of Kelly log betting: a tale of five investors, Tech Report UBC
3AbstractThe Kelly or capital growth criteria maximizes the expected logarithm as its utility function period by period.It has many desirable properties such as being myopic in that today’s optimal decision does not depend upon yesterday’s or tomorrow’s data,it asymptotically maximizes long run wealth almost surely and it attains arbitrarily large wealth goals faster than any other strategy.Also in an economy with one log bettor and all other essentially different strategy wagers, the log bettor will eventually get all the economy’s wealth.The drawback of log with its essentially zero Arrow-Pratt absolute risk aversion is that in the short run it is the most risky utility function one would ever consider.Since there is essentially no risk aversion, the wagers it suggests are very large and typically undiversified.Simulations show that log bettors have much more final wealth most of the time than those using other strategies but can essentially go bankrupt a small percentage of the time, even facing very favorable investment choices.One way to modify the growth-security profile is to use either ad hoc or scientifically computed fractional Kelly strategies that blend the log optimal portfolio with cash. to keep one above the highest possible wealth path with high probability or to risk adjust the wealth with convex penalties for being below the path
4Abstract (cont’d)For log normally distributed assets this simply means using a negative power utility function whose risk aversion coefficient is 1:1 determined by the fraction and vice versa.For other asset returns this is an approximate solution.Thus one moves the risk aversion away from zero to a higher level.This results in a smoother wealth path but usually has less growth.This talk is a review of the good and bad properties of the Kelly and fractional Kelly strategies and a discussion of their use in practice by great investors and speculators most of whom have become centi-millionaires or billionaires by isolating profitable anomalies and betting on them well with these strategies.The latter include Bill Bentor the Hong Kong racing guru, Ed Thorp , the inventor of blackjack card counting who compiled one of the finest hedge fund records.Both of these gamblers had very smooth, low variance wealth paths.Additionally legendary investors such as John Maynard Keynes (0.8 Kelly) running the King’s College Cambridge endowment, George Soros (? Kelly) running the Quantum funds and Warren Buffett (full Kelly) running Berkshire Hathaway had similarly good results but had much more variable wealth paths.The difference seems to be in the choice of fraction and other risk control measures that relate to true diversification and position size relative to liquid assets under management.
5Success in investments has two key pillars: devising a strategy with positive expectation andbetting the right amount to balance growth of one’s fortune against the risk of losses.This talk discusses the Kelly or capital growth log utility criteria for investing.A strategy which has wonderful asymptotic long run propertiesthe log bettor will dominate other strategies with probability one andaccumulate unbounded amount more wealth.But in the short run the strategy can be very risky since it has very low Arrow-Pratt risk aversion.
6Fractional Kelly strategies provide more security but with less growth. Examples from blackjack, horseracing, lotteries and futures trading illustrate the theory and its use in practice.I have been fortunate to work/consult with seven individuals who turned a humble beginning with essentially zero wealth into hundreds of millions (at least five are billionaires) using security market imperfections and anomalies in racing, futures trading and options mispricings.Once they reach million, then often log --> linear: bet on anything with a “positive expectation” as long as you diversify and move their wealth into the best hedge and alternative investment fundsAll of them used Kelly or fractional Kelly betting strategies.
7Some points to learn from this research Means are by far the most important aspect of any portfolio problem.You must have the mean right to have good performance.If you have the mean right and do not overbet you should do well.In levered bets, it’s the left tail that can lead to trouble so you must not overbet or you can have a large disaster occurring without warning.Behavioral and other anomalies can yield strategies that have positive means.These biases yield ideas that yield profitable positive mean strategies in racing, sports betting and options markets.The capital growth or Kelly criterion strategy yields the most wealth in the long run and dominates all other essentially different strategies.But in the short run, the expected log criterion with its essentially zero Arrow-Pratt risk aversion index is very risky and can have substantial losses.The most you should ever bet is the log optimal amount; betting more is suboptimal and betting double yields a zero growth rate.Negative power utility, which blends cash with the expected log maximizing portfolio provides more security but has less long run growth.These fractional Kelly strategies are attractive for many investment situations; determination of what fraction to use depends on constrained optimization models.
8Growth versus Security: Tradeoffs in Dynamic Investment Analysis One is faced with a sequence of investments in periods 1, …, n some favorable, some unfavorableGiven an initial fortune, how should one invest over time to have long-run growth of their fortune while at the same time maintaining its security?Develop computational schemes so that the investor can have a desired growth and security tradeoff.Find simple operational policies that achieve these tradeoffsTransactions costs are crucial in practice so stochastic programming is neededUse results to analyze favorable investment situations
9POLAR APPROACHES , Ziemba (2003) Markowitz (1976), Hausch, Ziemba and Rubinstein (1981, 1985), Luenberger (1993) and others below
11If you bet on a horse, that’s gambling If you bet on a horse, that’s gambling. If you bet you can make three spades, that’s entertainment. If you bet cotton will go up three points, that’s business. See the difference? Blackie Sherrod
12Games: favorable or unfavorable Blend growth versus security to your risk tolerance and the situation at hand
13Effect of data input errors on portfolio performance
14It’s the means that are the most important for investment success
15Mean Percentage Cash Equivalent Loss Due to Errors in Inputs Pension funds mix, RA=4 (Kallberg-Ziemba, 1983, Management Science)Conclusion: spend your money getting good mean estimates and use historical variances and covariancesReference: Chopra and Ziemba (1993), Journal of Portfolio Management, reprinted in Ziemba-Mulvey (1998) Worldwide asset and liability management, Cambridge University PressResults similar in period 1 of multiple period models and the sensitivity is especially high in continuous time models. See examples in AIMR, 2003.
16The results here apply to essentially all models The results here apply to essentially all models. You must get the means right to win!Optimal asset weights at stage 1 for varying levels of US equity means in a multiperiod stochastic programming pension fund model for Siemens Austria: see Geyer and Ziemba (2007, Operations Research)’s euro equities (.9US)US bonds 7.2, =11.3Euro bonds 6.8, =3.7
17Assuming the the mean return for US stocks is equal to the long run mean of 12% as estimated by Dimson et al. (2002, 2006) --> the model yields an optimal weight for equities of 100%.A mean return for US stocks of 9% --> < 30% optimal weight for equities.This is in a five period ten year stochastic programming model.The sensitivity to the mean is much less in periods 2, …, T
18Asset proportions: not practical: bonds vs stocks vs T-bill futures
19The Symmetric Downside-Risk Sharpe Ratio The Sharpe ratio is a very useful measure of investment performance.However, it is based on mean-variance theory and thus is basically valid only for quadratic preferences or normal distributions.Hence skewed investment returns can lead to misleading conclusions.This is especially true for superior investors such as Warren Buffett and others with a large number of high returns.Many of these superior investors use capital growth wagering ideas to implement their strategies which leads to higher growth rates but also higher variability of wealth.A simple modification of the Sharpe ratio to assume that the upside deviation is identical to the downside risk provides a useful modification that gives more realistic results.
21Using a modified Sharpe ratio that does not penalize gains Summary over funds of negative observations and arithmetic and geometric means
22The symmetric downside Sharpe ratio performance measure we want to determine if Warren Buffett really is a better investor than the rather good but lesser funds mentioned here, especially the Ford Foundation and the Harvard endowment, in some fair way.The idea is presented in a Figure below where we have plotted the Berkshire Hathaway and Ford Foundation monthly returns as a histogram and show the losing months and the winning months in a smooth curve. We want to penalize Warren for losing but not for winning. So define the downside risk asThis is the downside variance measured from zero, not the mean, so it is more precisely the downside risk.To get the total variance we use twice the downside variance
23The wealth levels from December 1985 to April 2000 for the Windsor Fund of George Neff, the Ford Foundation, the Tiger Fund of Julian Robertson, the Quantum Fund of George Soros and Berkshire Hathaway, the fund run by Warren Buffett, as well as the S&P500 total return index.
24Ford Foundation and Harvard Investment Corporation Returns, quarterly data, June 1977 to March 2000
25Comparison of ordinary and symmetric downside Sharpe yearly performance measures Only Buffett improves but he still does not beat the Ford Foundation - and Harvard is also better than Buffett but not Ford with the quarterly dataWhy? Tails still too fatThorp (2006) shows that Buffett is essentially a full Kelly bettor.
26Berkshire Hathaway versus Ford Foundation, monthly returns distribution, January 1977 to April 2000
27Return distributions of all the funds, quarterly returns distribution, December 1985 to March 2000
32Log UtilityIn the theory of optimal investment over time, it is not quadratic (the utility function behind the Sharpe ratio) but log that yields the most long term growth.But the elegant results on the Kelly (1956) criterion, as it is known in the gambling literature and the capital growth theory as it is known in the investments literature, see the survey by Hakansson and Ziemba (1995) and MacLean and Ziemba (2006), that were proved rigorously by Breiman (1961) and generalized by Algoet and Cover (1988) are long run asymptotic results.However, the Arrow-Pratt absolute risk aversion of the log utility criterion is essentially zero, where u is the utility function of wealth w,, and primes denote differentiation.The Arrow-Pratt risk aversion index.is essentially zero, where u is the utility function of wealth w, and primes denote differentiation.Hence, in the short run, log can be an exceedingly risky utility function with wide swings in wealth values.
33Long run exponential growth is equivalent to maximizing the expected log of one period’s returns
34Thus the criterion of maximizing the long run exponential rate of asset growth is equivalent to maximizing the one period expected logarithm of wealth. So an optimal policy is myopic.Max G(f) = p log (1+f) + q log (1-f) f* = p-qThe optimal fraction to bet is the edge p-q
35Slew O’ Gold, 1984 Breeders Cup Classic f*=64% for place/show; suggests fractional Kelly.
39Kelly and half Kelly medium time simulations: Ziemba-Hausch (1986) These were independent
40The good, the bad and the ugly 166 times the wealth is more than 100 times initial wealth fail with full Kelly but only once with half KellyBut probability of being ahead is higher with half Kelly, 87% vs 95.4%Min wealth is 18 and only 145 with half Kelly700 bets all independent with a 14% edge, result you still lose over 98% of your fortune with bad scenariosWith half Kelly, lose half of wealth only 1% of the time but 8.40% with full Kelly
41Kentucky DerbyUse inefficient market system in Hausch, Ziemba, Rubinstein (1981) and Ziemba-Hausch booksPlace/show wagers made when prices off sufficiently and EX≥ 1.10w0 = $ years 72 wagers with 45 (62.5%) successful
42Typical wealth level histories with one scenario (the actual results) from place and show betting (Dr Z system) on the Kentucky Derby, with Kelly, half Kelly and betting on the favorite strategies
43Should you ever be above 0.02 that is positive power utility like OverbettingProbability of doubling and quadrupling before halving and relative growth rates versus fraction of wealth wagered for Blackjack (2% advantage, p=0.51 and q=0.49Should you ever be above 0.02 that is positive power utility likeI think its dominated!Betting more than the Kelly bet is non-optimal as risk increases and growth decreases; betting double the Kelly leads to a growth rate of zero plus the riskfree asset.LTCM was at this level or more, see AIMR, 2003.
44Growth Rates Versus Probability of Doubling Before Halving for Blackjack
45Fractional Kelly and negative power utility u(w) =-w <0 0 u logf=1/(1- ) = fraction (Kelly) in log optimal portfolio, rest in cash=0 f=1 full Kelly=-1 f=1/2 1/2 Kelly=--3 f=1/4 1/4 Kelly futures trading down hereThis is exact with log normaility and approximate otherwise.
47Commodity Trading: Turn of Year Effect Small cap stocks have outperformed large cap stocks in January on a regular basis since 1926Average excess returns of smallest minus largest decile of US stocks, , Source: Ibbotson Associates
48sellbuyFutures play with anticipation, mid December to mid January, this is a typical year in the mid 90s, Value Line versus S&P,
49Turn of the year effectProbability of reaching $10 million before ruin for Kelly, half Kelly and quarter Kelly strategiesRelative growth rate and probability of doubling, tripling or tenfolding before halving for various Kelly strategies
50Turn of the year effect, recent developments Futures markets - much more violentRussell has more volume than Value LineEffect moved into DecemberTextbooks and finance experts say effect is not thereGraphs in Hensel-Ziemba paper in Keim-Ziemba (2000) Worldwide security market imperfections, Cambridge University Press.Doing this trade is like driving a dynamite truck smoking a cigar. You do it carefully.Rendon-Ziemba (2005) update to 2005 turn of the year Value Line/S&P500 and Russell 2000/S&P500 spread trades
51Unpopular numbers in the Canadian 6/49, 1984, 1986, and 1996 Lotto
53Probability of doubling, quadrupling and tenfolding before halving, Lotto 6/49 Case ACase B
54Probability of reaching the goal of $10 million before falling to $25,000 with various initial wealth levels for Kelly, 1/2 Kelly and 1/4 Kelly wagering strategiesThe downside of the analysis is that the expected time to win a lot is in the millions of years.
55The InvestorsTom, I believe, is overbetting and dominated and will go bankruptHarriet has a limited degree of risk tolerance, fits well with lots of empirical Wall St equity premium data
68Horseracingmarket in miniaturefundamental and technical systemsreturns and odds are determined by1) participants -- like stock market, unlike roulette2) transaction costs -- track take (17%), breakage;rebates now plus Betfair (long short)bet to1) win -- must be 1st2) place -- must be 1st or 2nd3) show -- must be 1st, 2nd or 3rd
69Place market in horseracing Inefficiencies are possible since:1) more complex wager2) prob(horse places) > prob(horse wins) ==> favorites may be good betsTo investigate place bets we need:1) determine place payoffs 2) their likelihood 3) expected place payoffs4) betting strategy, if expected payoffs are positiveBettors do not like place and show bets.
70The IdeaUse data in a simple market (win) to generate probabilities of outcomesThen use those in a complex market (place and show) to find positive expectation betsThen bet on them following the capital growth theory to maximize long run wealth
71Effect of transactions costs, calculation of optimal place and show Kelly bets Non concave program but it seems to converge.In practice, adjust q’s to replicate biases.
73Use in a calculatorWhat we do in the system is to reduce the non-convex log optimization problem down to four numbers: Wi,, W, and Si, S or Pi, P,Thousands of race results regress the expected value and the optimal Kelly bet as a function of these four variables.Hence, you just find horses where the relative amount bet to place or show is below the bet in the win pool.The calculator tells you when the expected value is say 1.10 or better and calculates the optimal Kelly bet.So this can be done in say 15 seconds.
76Expected value approximation equations Expected value (and optimal wager) are functions of only four numbers - the totals and the horse in question.These equations approximate the full optimized optimal growth model.Solving the complex NLP: too much work and too much data for most people.This is used in the calculators, and Hausch-Ziemba (1985, Management Science), differing track take, etc.
79Simulations in 2004-5 Real results April 2005-March 2006 Up ~ 36,000 ~ 2% on bets ~ 1.5 M,System -7%, rebate ~ 9%, edge ~ +2%
80Kelly fractions and path achievement Calculating the optimal Kelly fractionTo stay above a wealth path using a Kelly strategy is very difficultKelly fractions and path achievementthe more attractive the investment opportunity,the larger the bet size andhence the larger is the chance of falling below the path.MSZZ using a continuous time lognormally distributed asset model calculate that function to stay above a path at various points in time to stay with a high exogenously specified value at risk probability.Convex case like Geyer-Ziemba (2007) Siemens Vienna pension model - can do on a computer; will develop the math
81The planning horizon is T=3, with 64 scenarios each with probability 1/64
82With initial wealth W(1)=1, the value at risk is a With initial wealth W(1)=1, the value at risk is a. The optimal investment decisions and optimal growth rate for a, the secured average annual growth rate and 1-a, the security level are shown in the table.
89Some properties of the Capital Growth Theory (cont’d)
90Some properties of the Capital Growth Theory (cont’d)
91Some properties of the Capital Growth Theory (cont’d)
92Some properties of the Capital Growth Theory (cont’d)
93Some properties of the Capital Growth Theory (cont’d)
94ReferencesEssentially all of the material in this talk is in the following books plus the papers handed outZiemba, The Stochastic Programming Approach to Asset Liability Management, AIMR, 2003Ziemba-Hausch, Dr Z’s Beat the Racetrack, William Morrow, 1987 (has UK betting system)Hausch-Lo-Ziemba, Efficiency of Racetrack Betting Systems, Academic Press, Classic new and reprinted articles, bible for Hong Kong professional betting teams. Originals sell for huge prices as high as $12,000 I am told, I sold one for $1400 last week.Ziemba-Vickson, Stochastic Optimization Models in Finance, Academic Press, Classic articles, new articles, huge collection of portfolio theory, problems. Reprinted by World Scientific, Singapore, 2006.Ziemba et al, 6/49 Lotto Guidebook, 1986Ziemba-Hausch, Betting at the Racetrack, 1986, exotic bet pricingBooks all available, for information.