Blog - Investing Notes
March 24, 2006 - Stock Trading Wisdom in the Crowd of AI Software?
Reader Eric poses the following question:
"Have you evaluated and compared TradingSolutions 3.1 versus DeepInsight in terms of accuracy and usability?"
From the supplier's web site: "TradingSolutions is a comprehensive technical analysis software package that helps you make better trading decisions by combining traditional technical analysis with state-of-the-art neural network technologies." And: "DeepInsight combines quantitative analysis with Artificial Intelligence to analyze trading patterns and market data in a great depth, and therefore, predict and catch trends in early stage for individual stocks, ETFs, mutual funds and market indices." We do not do product evaluations, but can offer some research-grounded observations on acquiring and using artificial intelligence (AI - neural network, expert system, genetic algorithms...) software to guide stock selection and trading performance. Specifically:
There are a few prerequisites to a return on an investment in AI software:
1. You must believe that markets are significantly and systematically inefficient such that examination of past information (prices, volumes, macroeconomic data...) supports forecasting of future returns. Are there really any patterns for the AI software to learn? There is plenty of research to support a belief that some patterns exist (see the summary of the emergence of behavioral finance).
2. You must have reliable methods to find the inefficiencies. For example, for short-term trading, recall the formal research on 39,832 simple trading rules. There is much conventional wisdom that does not work. Does the AI software provide rules that work? With lots of compute power, it's easy to find arcane "rules" that work in the past but not in the future.
3. You must understand how to trade in a statistical framework, with enough trades (capital) to let the probabilities work for you. You will have lots of losers as well as winners. Transaction costs will add up. Practice and speed appear to be critical success factors. To quote an expert: "Neural nets won't make a bad trader good, they'll make a good trader better." Note that, from a statistical perspective, product testimonials are next to worthless because testimonial samples are too small and undoubtedly selection-biased.
There is not a lot of very recent research specifically testing AI for stock trading. That which we could find generally concludes that AI can generate abnormal returns. Here are two examples:
"Financial Prediction and Trading Strategies Using Neurofuzzy Approaches"
The authors of these papers are engineers and scientists, not financial market experts or professional traders. We infer that working with AI software may therefore be challenging, and that conclusions from the research may reflect material lack of experience with real-life stock trading.
The performance of trading systems at Collective2 is instructive. Some of these systems are, at least nominally, AI-based.
There is not enough information on their web sites to distinguish clearly between TradingSolutions 3.1 and DeepInsight. The suppliers of the former appear to be more sophisticated from a marketing/presentation perspective, and they offer a working evaluation copy, but their product is more expensive.
See also our blog
entry of 9/15/05 for reactions to the disclaimer provided by the
suppliers of VantagePoint Trading Software. One reaction there, that
it is apparently easier for the supplier to make money by selling the
software than by trading with it, also applies here. The disclaimers
for TradingSolutions 3.1 versus DeepInsight are much shorter and less
revealing.
