OVERVIEW

1.     Not all Artificial Intelligence is created equal, but that is not nearly as important as understanding the AI system's inputs.

2.     In designing a smart stock selection system utilizing Artificial Intelligence, anyone can use inputs like price, volume, dividend yield, market cap, etc. “the current analytical metrics”. CONCERT AI goes far beyond this.  CONCERT AI uses small packets of powerful quantitative indicators as inputs to its Artificial Intelligence:

·       The MTI (“Market Temperature Indicator”) which is analogous to temperature in the Maxwell Boltzman physics model. (for definition, click here)

·       Quantified Japanese Candlesticks (for definition, click here)

·       A processing layer to normalize the data to allow equal weighting of inputs.

·       The LSI (long strength indicator) and SSI (short strength indicator), which includes commodity data such as the price of oil, interest rates, price of gold, real estate, etc.

 

Neural Network Architecture in CONCERT AI

Neural networks are mathematical algorithms characterized by hundreds of interconnected modules (the node functions, e.g., sigmoid) and empirical weightings along the interconnection lines. They mimic the structure of brain neurons to understand/model real-world data.

Neural networks comprise about a third of our system. CONCERT AI uses several dozen neural networks, all modest in size (no more than 11 independent variables) that represent several 'families', each with slightly differing objectives. These neural networks were last trained in 2007. They utilize a 'rich diet' of independent variables that include proprietary indicators, functions of prices, volumes, indices and commodity prices. The system avoids 'curve-fitting' by using 16 years of market data with only a couple of hundred node weights.

Another major component of our system are the 1000 Grail System models. They are individual stock-oriented Expert Systems with an if-then-else structure and a corresponding set of 'empirical parameters'.

Functionally, the combination of Neural Networks and the Grail system forms a superior AI capability that is less susceptible than Neural Networks alone to 'curve-fitting'. With only 69 parameters available to characterize the 16-year performance of a single equity, there are insufficient parameters to 'mimic' the data. Instead, there are enough parameters to 'predict' what how each equity is likely to perform. The goal is to use rules which describe the 'underlying' forces that influence the price behavior of each stock, with adequate parameters to model behavior for decades to come.

Note also that since these models 'learn' the patterns of a stock for our full 16-year training period, and their training criteria regard all time intervals as equally valid (the accuracy in 2002-2003 is just as important as the accuracy in 2017-2018), their empirical parameters are slow to change.

CONCERT AI utilizes multiple model evolution 'layers'. The equity signals for our current portfolios come directly from the Grail System models which utilize our neural network families and their own associated 69-parameter sets, but the Portfolio Rules incorporate information from the Preprocessor, the L/7 holdings, which are 'historical', our MFTI-B (long-term) signal, and our current ultimate layer, System6, which provides next day market strength assessment focused on the S&P. Our CONCERT AI portfolios succeed because all of these elements are balanced to maximize gain and minimize risk.

CONCERT AI a composite entity, with small, specialized Neural Network modules embedded within, and connected by, Expert Systems. The entire system continually learns about market behavior. In our current system the neural networks were last trained in 2007, but all of the Expert System 'glue' that binds everything together undergoes varying degrees of ongoing learning.

CONCERT AI is therefore not the kind of Neural Network 'AI' system that could be assembled in days, weeks, or months. Our composite system is the culmination of 16 years of algorithmic research and evolution.

 

CONCERT AI SYSTEM ARCHITECTURE

The following diagram illustrates the overall architecture of the CONCERT AI Neural Net system, and how a large number of securities are filtered for selection to optimize return while minimizing drawdown.

funnel1.jpg

 

CONCERT AI systematically reduces the number of potential investment candidates by a highly structured process. This first produces a list of 1672 securities that demonstrate a tendency to repeat pattern outcomes. The system subsequently filters down to 1000 candidates that undergo additional rigorous neural network screening. 69 parameter coefficients identify the candidates with the highest predictive accuracy and expected Return/Maximum Drawdown ratio.  In the final stages of filtering, the system identifies patterns that demonstrate the highest probability of success, often exceeding 80%, and recommends investments.  If attractive options do not exist, the system selects cash as the desired investment.

CONCERT AI will continue to hold investment positions when the patterns that produced “buy” signals continue to exist, selling securities when the recognized patterns fall below expectations or turn negative.  The system uses additional rules to discipline the purchase and sale of securities. These rules include market temperature indicators, relative strength measurements and risk adjusted returns.

CONCERT AI generally competes favorably against recognized indices that must stay fully invested through all investment cycles.  These benchmarks never invest in cash to avoid down markets and only make minor changes in quarterly allocations based on capitalization weighting. CONCERT AI makes no attempt to replicate the indices or minimize tracking error, but strives to outperform by taking advantage of:

  • Investments showing the highest likelihood of future profits.
  • Cash as a safe harbor where investment may be diverted to avoid perceived or real market turmoil and declining prices.

Performance Is the paramount CONCERT AI consideration.