3 ms·

> Update: Added a Monte Carlo mode which lets you play with data that is randomly generated from the daily returns of the S&P500. The probability of a daily ret

> Update: Added a Monte Carlo mode which lets you play with data that is randomly generated from the daily returns of the S&P500. The probability of a daily return being picked is the same probability/frequency that it occurred in the last 68 years. This mode is rigged. Any proposal for market timing requires correlated returns. "Technical" traders infer short-term trends form patterns like the shave-and-a-haircut and lovely-lady-humps that are ultimately based on a theory of market psychology, and "fundamental" traders usually make structural observations like price-to-earnings ratios being mean-reverting. Both of these hypotheses are excluded by construction when market results are constructed by a random draw of daily returns.

grey-area 7y ago
Yes exactly. The market may resemble a random walk in the short term, over days, but it is driven by economic forecasts and company results in the long term (the long term is also more than the 3 year timing given). Absent any of that information, obviously you can’t guess how the stocks will move (i.e. time the market), except that they will in aggregate move gradually upwards in price as all currencies aim for inflation. With that information, sometimes the way the market will move over a time period of years is clear. Clearly the consensus moves in a panicky way with the latest news, and constantly undershoots or overshoots real returns, but it is based on real returns in the long run.
jcfrei 7y ago
They could easily create a better simulation with a moving block bootstrap method. That is by sampling consecutive returns (for example for a whole week) and create a new time series with them.
- hogFeast 7y ago
  Yep, I have a tool that is similar to this and this is what I did. Unfortunately, if you are working with daily returns you need consecutive returns for way more than a week. You have correlated returns, in particular there are Friday-next Monday correlations that are important in the tails. You also have volatility clustering/asymmetries over daily periods (i.e. high volatility tends to be followed by higher volatility and the volatility responds differently to up vs down moves) and this tends to last way longer than a week. It is very tricky stuff. In the real world, you will often find managers grouping based on their knowledge (i.e. X-Y was the 2008 crisis) and testing their portfolios against that. It is rather unscientific but it works (i.e. in this case, you might do something like a Markov model with a transition matrix of the daily probability of moving between volatility states, and then sample longer blocks from groups based on the state). A simpler option (what I did) is to just look at yearly returns, and sample across countries (just using the US is horrible cherrypicking).
TomMckenny 7y ago
Yes, it simulates the hypothesis that future markets are unknowable. Which would indeed be begging the question. Unless of course you find that whenever you apply technical trading to a Monte Carlo market, you get the same results as applied to real market history. Then you show that the signals technical trading uses are not actually predictive. At any rate, if the market trends upward and trading is close to random (regardless of what the trader believes) then being out of the market occasionally will always be bad statistically. If there are vast numbers of traders all actually acting randomly then some will out preform others and some tiny number might out preform the market. In fact, if you have the data, you can plot the performance of every trader against the market average and actually measure how far from random the average trader is. Maybe price to earnings ratio "should" relate to stock price. But in actuality stock price is determined by what people are willing to pay for it obviously. To the degree that people are buying and selling _not_ based on P/E, say for example they are doing so somewhat randomly, then the P/E won't reflect accurately the stock price: it won't be predictive. These hulking paragraphs is what the demo is basically saying.
valuearb 7y ago
Fundamental investing doesn’t require mean reversion. And technical patterns aren’t predictive.
rmrfstar 7y ago
Actually, this is not correct. Any proposal for market timing requires an observable x[t] that is correlated to R[t+d] for d>0. The returns may themselves lack any auto-correlation. If market returns are auto-correlated in some way, then x[t] could be a function of R[s] for s<=t. In the context of this game, x[t] would be some information about the internal state of the RNG at time t.

grey-area 7y ago

Yes exactly. The market may resemble a random walk in the short term, over days, but it is driven by economic forecasts and company results in the long term (the long term is also more than the 3 year timing given). Absent any of that information, obviously you can’t guess how the stocks will move (i.e. time the market), except that they will in aggregate move gradually upwards in price as all currencies aim for inflation. With that information, sometimes the way the market will move over a time period of years is clear. Clearly the consensus moves in a panicky way with the latest news, and constantly undershoots or overshoots real returns, but it is based on real returns in the long run.

jcfrei 7y ago

They could easily create a better simulation with a moving block bootstrap method. That is by sampling consecutive returns (for example for a whole week) and create a new time series with them.

hogFeast 7y ago
Yep, I have a tool that is similar to this and this is what I did. Unfortunately, if you are working with daily returns you need consecutive returns for way more than a week. You have correlated returns, in particular there are Friday-next Monday correlations that are important in the tails. You also have volatility clustering/asymmetries over daily periods (i.e. high volatility tends to be followed by higher volatility and the volatility responds differently to up vs down moves) and this tends to last way longer than a week. It is very tricky stuff. In the real world, you will often find managers grouping based on their knowledge (i.e. X-Y was the 2008 crisis) and testing their portfolios against that. It is rather unscientific but it works (i.e. in this case, you might do something like a Markov model with a transition matrix of the daily probability of moving between volatility states, and then sample longer blocks from groups based on the state). A simpler option (what I did) is to just look at yearly returns, and sample across countries (just using the US is horrible cherrypicking).

TomMckenny 7y ago

Yes, it simulates the hypothesis that future markets are unknowable. Which would indeed be begging the question. Unless of course you find that whenever you apply technical trading to a Monte Carlo market, you get the same results as applied to real market history. Then you show that the signals technical trading uses are not actually predictive. At any rate, if the market trends upward and trading is close to random (regardless of what the trader believes) then being out of the market occasionally will always be bad statistically. If there are vast numbers of traders all actually acting randomly then some will out preform others and some tiny number might out preform the market. In fact, if you have the data, you can plot the performance of every trader against the market average and actually measure how far from random the average trader is. Maybe price to earnings ratio "should" relate to stock price. But in actuality stock price is determined by what people are willing to pay for it obviously. To the degree that people are buying and selling _not_ based on P/E, say for example they are doing so somewhat randomly, then the P/E won't reflect accurately the stock price: it won't be predictive. These hulking paragraphs is what the demo is basically saying.

valuearb 7y ago

Fundamental investing doesn’t require mean reversion. And technical patterns aren’t predictive.

rmrfstar 7y ago

Actually, this is not correct. Any proposal for market timing requires an observable x[t] that is correlated to R[t+d] for d>0. The returns may themselves lack any auto-correlation. If market returns are auto-correlated in some way, then x[t] could be a function of R[s] for s<=t. In the context of this game, x[t] would be some information about the internal state of the RNG at time t.