In this thread I will be happy to discuss and learn more about ACE and it's implications for practical trading systems.A good place to start is: http://www.econ.iastate.edu/tesfatsi/ab ... tm.Regards

QuoteOriginally posted by: yabbadabbaIn this thread I will be happy to discuss and learn more about ACE and it's implications for practical trading systems.A good place to start is: http://www.econ.iastate.edu/tesfatsi/ab ... .RegardsIs agent technology difficult? I see lots of research being done here, but no software products.

QuoteOriginally posted by: yabbadabbaIn this thread I will be happy to discuss and learn more about ACE and it's implications for practical trading systems.why do you need to simulate fake markets when you can study REAL stock market data?

I have done a ton of economic simulations. I find they are slightly useful in thinking about how something works, but not in predictions.I find I can't simulate something unless I can already understand and it better than any simulation could explain or reproduce. So when trying to understand associations between inputs, I find that my mind works better. And once discovered, the associations are very simple, so that a simple ruleset can be used to trade them.I can also conceive of and simulate changes in the associations between inputs. But I cannot use those same prices to predict changes in their associations. I can only observe the changes in real time.In short, most information which maniests in the output of simulations is invented, and not observable.

QuoteOriginally posted by: farmerI have done a ton of economic simulations. I find they are slightly useful in thinking about how something worksYou can think about how something works without doing any simulation. Einstein did not do simulations to discover special/general relativity. He just thought it through. Simulations are for lazy hacks who want to substitute coding for serious brain work.

I do agree with Farmer that ACE is more valuable for understanding models of participant behavior and market dynamics overall, rather than direct predictions in specific market conditions. Yet I find ACE valuable in two ways (in addition to what Farmer suggested):1. ACE represents a step deeper into the causal network of economic dynamics compared to MC or stochastic calculus techniques that assume a distribution and derive results from there. Rather than asserting that the price time series is IID Gaussian or GARCH or some jump diffusion model, ACE tries to simulate the underlying participant behavior (perhaps pumped by exogenous events and endogenous feedback cycles) to compute price action. If one has a model for why people trade/speculate/invest, then one can use ACE to test that model. If the resulting simulated system doesn't "behave like" real markets, then it leads to deeper understanding of the discrepancy between simple models of participant behavior and that real behavior. I've not used the simulations to estimate real market parameters, but I suppose one could do that.2. ACE is a data augmenter. Even though we have tons of real market data, it really is finite and easily overfitted. It's easy to run out of data when playing-slice-and-dice with in-sample/out-of-sample splits and exploring alternative statistical models, parameter spaces, etc. Synthetic datasets let one develop trading systems without "using up" market data. This includes intentionally creating datasets that do or do not contain trading opportunities to help gauge the potential for overfitting.Overall ACE is a way of thinking about the markets, rather than predicting the markets. It provides a repeatable sandbox for doing experiments in market behavior that are impossible in real markets.

agent simulation is for hacks who use computers to hide their lack of ability to think mathematically. CAPM has agents too, but the inventors of CAPm did not need to simulation anything. They just thought it out and proved it mathematically. Clean and penetrating.

The reason for simulation is that the analysis is too complex for simple closed form solutions. Can you tell the the CAPM formula for agents with dynamic heterogenous beliefs and preferences that are correlated over time?

QuoteOriginally posted by: gardener3 Can you tell the the CAPM formula for agents with dynamic heterogenous beliefs and preferences that are correlated over time?No, but you cannot derive any formula using simulations either. If something is "too complicated" to analyze using rigorous analytical analysis, you won't learn much about it using simulations either. Simulations are ONLY useful when you have a closed form solution and you want to study some perturbations of the closed form solution. All other simulations are black boxes and confuse rather than clarify.

QuoteOriginally posted by: farmerI have done a ton of economic simulations. I find they are slightly useful in thinking about how something works, but not in predictions.I find I can't simulate something unless I can already understand and it better than any simulation could explain or reproduce. So when trying to understand associations between inputs, I find that my mind works better. And once discovered, the associations are very simple, so that a simple ruleset can be used to trade them.I can also conceive of and simulate changes in the associations between inputs. But I cannot use those same prices to predict changes in their associations. I can only observe the changes in real time.In short, most information which maniests in the output of simulations is invented, and not observable.Pretty spot on with my own experience. I do find that 3-4 times a year there are are predictive effects. You rarely need to model or simulation once you have the causal dynamics worked out. In fact, this forms the basis of my trading.

Oops

Multiple message being posted by yours truly - a lag between the button and appearance on the page does not mean the action has not been registered.

@KackToodlesSimulations are good finding new solutions to problems where not all possible solutions can be iterated, hence don't have a closed formula. Simulations are tools very widely used today in Physics, Chemistry, social sciences. Saying that simulations is only for stupid people is hilarious.

QuoteOriginally posted by: KackToodlesagent simulation is for hacks who use computers to hide their lack of ability to think mathematically. CAPM has agents too, but the inventors of CAPm did not need to simulation anything. They just thought it out and proved it mathematically. Clean and penetrating. But mathematical proof is a type of simulation. Instead of writing code to numerically express the behavior and structure of a system, the mathematically-inclined write a bunch of axioms, lemmas, and theorems to symbolically express the behavior and structure of a system. But both numerical and analytic models of a system are simulations in the sense that they are not the thing itself, but are representations of the thing. In fact they are probably freely interconvertable. One can certainly create a numerical version of a symbolic model and I'd bet that one could create a C-to-Mathematica compiler and convert any C program numerical simulation into a "closed form" representation.Both symbolic (closed-form) and numerical simulation suffer from the same problem -- they use reduced-complexity representations of the actual system. Both need some validation process that tests both components of the simulation (whether symbolic or numerical) and the results of the whole of the model against empirical results.I'm not knocking closed-form symbolic simulations -- analytic structure can reveal some incredibly deep insights. Yet, symbolic simulations do tend to suffer from excessive simplification -- we're forced to apply axioms that make the math tractable rather than choose axioms that have empirical support. Analytic models often ignore known structure by claiming its effects disappear through averaging or when the system moves equilibrium. Numerical simulations permit a more complex model but probably yield a shallower set of insights.If you don't use both types of simulations, there's a much higher chance that you will be fooled by either randomness or the seduction of the elegance of your closed-form (but erroneous) models.

Brilliant post Traden4Alpha! There is indeed a very close relationship between 'numerical' mathematics and programs. Each expression and each proof can be seen as a program. An expression is a product of some formal language. A proof is a list of transformations, subject to logical rules. This is why logic is at the heart of both mathematics and computer science.Classical economics takes closed formulas and provides numerical solutions. I studied economics and I found quite dull. Most economists don't bother about testing the theories using economectrics or empirical tests in general. Instead there are extremely fond of toyproblems, weird assumptions and simple closed systems. Most theories in economics are much to local to explain anything; most of the time they should be seen as mere ideas of how to think about a set of problems. Of course there are not to blame, because up to fairly recently the tools and the paradigm of complex systems was almost completely unheard of.-------One point to think about is IMHO if and how traders/quants use economical theory at all. My guess is they mostly don't, because it is too complex, too much work. But, on the other hand companies are often the driving force of innovation (think about computer chips e.g.), because they have much more resources (people&capital&knowledge). I think that traders have much more knowledge about price fluctuations and anomalies than any scientist. But if they don't share their knowledge, every trader has to start from zero. The main difference between a trader and a scientist is that the trader is maximizing his profit and therefore significance of very specific forecasts, whereas the scientist is maximizing the significance of a wide variety of forecasts.