Replace artificial intelligence people in Wall Street?

AQR Capital Management and Yale University, one of the most important questions about the financing world: Will AI discuss researchers and traders? Through 44 pages of theories and experimental consequences, under the title “The virtues of the complexity in the prediction forecast,” Brian Kelly, Simon Malaamoud and Kengang Zhao claim that the most complicated models strong for humans are better than simple models in terms of reaching yields. As Bloomberg News reported, the reaction came to them quickly, as at least six research articles were stabbed in what they concluded, and Kelly defended the conclusions in turn. There is an opportunity to clone you with artificial intelligence … Why don’t you try? How will the result be? I bet on Kelly and his colleagues. Theoretical arguments are technical, but the most important question is old and easy to understand. One of the ways to predict is to look for some of the most important indicators with clear cross -links with the thing you want to predict, and merge it in simple ways. For example, to predict the return of the stock market for the next month, you can look at the return of this month, interest rates, price ratios to profits and such variables. Reading the past or the extrapolation of the future is all treated as random noise to be ignored, and the problem of using many indicators lies, or integrating them in very complicated ways, on the SO calls “excessive alignment” as you get a model that fully explains the past, while the future does not explain at all. It has built a model that used the sound of the past to explain everything, but these noise relationships will not continue in the future. If your interest is based on understanding or interpretation, the simple approach above is the best solution. But if your interest is limited to prediction, there is another way: add each indicator that can be introduced to the model, and the technical term for this approach is the ‘kitchen zinc model’. Try every complex combination. If the shares whose symbol describes the letter (v) tend to rise on Rainy on Tuesday, it is part of your model. The idea is that even if the index does not have a prediction value, it does not harm your expectations; It will just add noise. You can post everything and then reduce the noise later, or to be abundant of the trade, so that the diversification eliminates this noise. Will it increase productive artificial intelligence immediately? The discussion raised by the research article is deeper than this explanation. Kelly and others are not included in all possible indicators in their model, but only 15 variables, each of which have 12 monthly value – ie 300 value in general – and they enter 12,000 laboratories that predict the return of the stock market for the next month. They do not use letters in stock symbols or weather on Tuesday. But their opponents do not argue in favor of the most basic models; On the contrary, they deny the virtues of complexity. A very similar booking house half a century ago in the context of the roulette game. In the early 1960s, Edorpe, a professor of math, who created the calculation of Black Jacques game cards, and built with him Claude Shannon, the father of the information theory, the first portable computer in the world to predict the roulette sessions. The previous regulations to win roulette depend on the schedule of the previous results to find numbers that look more than others. Many people have argued that the roulette wheels are the manufacturing courts, just as it is possible to overcome them. An proof of an idea that many people consider impossible is that the most important idea of ​​Thorpe lies in the fact that if the roulette wheels are designed with adequate accuracy, so that each number appears on the same frequency, it must be predictable. His initial work showed that roulette rotation consists of two phases. If the ball against the outer edge of the bowl – the lane of the ball – turns – turns the head of the wheel (the moving part of all numbers) in the opposite direction, the system is controlled by the simple physics of Newton. Knowing the velocity of the ball, the head of the wheel and friction transactions, it will facilitate the prediction of the number that will be under the ball when it comes out of the track and ends up at the head of the wheel. But as soon as the ball comes out of the track, it makes reflexes, rotation and move, its movement is chaotic and difficult to predict. Knowing only the number under the ball when it comes out of the track can determine a third of the wheel in which the ball sits in 40% of cases – which is more than enough to earn a bet. How do you stop your artificial intelligence you work? This has led to one of the basics of quantitative investors: the opportunity is to find the ability to predict things that others see randomly, and uncertainty in the things they consider inevitably. By the 1970s, the construction of a portable roulette computer and its success became a quantitative baptism ritual. Technology improvements have led to major improvements in accuracy and reliability. When I tried it in the middle of the last century, the field branched. The group, the physicists, focused their efforts on improved measurement devices. Use complicated equations to process relevant data with the causal models of physics. Looking for defects, I was prone to the other group, namely the statistics. We have used primitive versions of machine learning algorithms to utilize patterns. Not only did we want our goal to take advantage of the inevitable factors of adopting the perfect roulette wheel, but rather the patterns caused by the defects, as some number boxes are soft or more difficult than others, or that the wheel is not completely horizontal. We measured a lot more physicists, but with less accuracy in each, and we analyzed a lot of data that may not be related. The two groups had very similar arguments for the current arguments about the virtues of the complexity. The main advantage of physicists was some or non -existent training devices on individual wheels, because of their dependence on comprehensive physical law instead of individual wheel defects. Our benefit was low cost and higher forecast accuracy – especially in less accurate casinos that use cheaper wheels and comfortable maintenance – at the expense of the need for hours of calibration before betrayal is profitable. They stopped in vain from comparing artificial intelligence with people. I have been betting on the complexity instead of theory, as well as the prediction instead of understanding. I always felt that machine learning and artificial intelligence would replace analysts and traders (as well as managers, doctors, advocates, scientists and others). You will find successful artificial learning algorithms and their own patterns through the greatest possible data, instead of leading people to selecting relevant data and imposing theoretical restrictions on the answers in advance. But I am wrong sometimes, so you do not bet with all your money on the number elected by my roulette calculator.