Epimenides Investment Technology

Our development team brings decades of work modeling human understanding of language to an application that 1) simulates an analyst's query process for finding events of relevance to price forecasting and 2) augments it by an order of magnitude, through increased volume of data throughput and a rigorous treatment of the relevant qualitative factors.

Our Approach

The investment climate is a fast-changing place, and the investment management industry has done a poor job in forecasting that change. Rigid models built on past events, human bias in processing new ones -- these difficulties give rise to the long-running, and oft-lost, competition in The Wall Street Journal between the "pros" and the dartboard, in making investment decisions. EIT brings a new approach. EIT credibly addresses some of the problems inherent in current approaches, using computational techniques that have only now become possible. Epimenides Investment Technology (EIT) applies Dr. Ezra Black's work in Natural Language Processing (NLP/Black) to the world of financial market modeling, by constructing probabilistic models to yield advanced information for trading purposes. These models of financial processes incorporate both traditional predictive factors and new, text-based predictive factors which can only be obtained by computationally understanding huge volumes of news, patents, reports and other documents. As proof of concept, work has been done over the past year to adapt NLP/Black to the platinum market. The platinum market has the basic characteristics of most other markets, with the benefit of its being a somewhat less complicated market than most, with high degrees of inefficiency. Our strategy has been to use NLP/Black to simulate the platinum market expertise of our business development partner and platinum market expert, Aran Murphy.

NLP/Black: A Technical Summary

The basis of the approach taken, in NLP/Black, to finding all and only the instances of a large variety of expertly-defined "text events" in huge document sets, is this: We apply sophisticated and probabilistically-trained first-pass filters to identify a set of documents for intensive processing. These are then parsed with an engine based on what is often considered the world's furthest-developed computerized English grammar, in development now for 20 years and offering full coverage of English with very detailed meaning and structural analysis of any input sentence of English. Each parsed document is then analyzed by an XML postprocessor defined over individually parsed sentences and their sub-trees, as well as paragraphs and distinguished document portions such as Title and Dateline. The analysis attempts probabilistically to match expertly-defined templates for specific "text events" which have significance, when they occur, for the specific investment issues being modeled, e.g. the probable price of platinum on the London market over a window of 3 days, 5 days, etc.

Applications in Finance

Current market modeling efforts have yielded results typically little better than random. Flipping coins to make investment decisions can often be as effective as the more formal decision-making methods on offer. Traders and managers try, nevertheless, to do better than random. There are two approaches usually taken, oftentimes simultaneously, with little to no connection or bridge between the two methodologies: 1. Pure "black box" trading strategies that use supercomputers and sophisticated models that look for minute price discrepancies. These run on a purely numerical basis, using market price data input. A problem with this is that with "black box" strategies, their predictive capabilities over random are so small and short-lived that high leverage, high volumes, and immediate execution of trades are necessary to realize gains from such trades. Another problem is that computers fail to incorporate major external changes that human managers are capable of seeing, such as risk of war or policy changes. 2. "Fundamental Analysis", which is the process by which seasoned investment managers ply their assumptions of causal relations and make decisions on the basis of their experience in the market and their human understanding of what they consider relevant factors. This approach is more fluid, better for adjusting to rapidly changing market contexts, and allows non-numerical feed information to be incorporated into a model. However, the "fundamental analysis" approach is too often riddled with untestable assumptions and lacks overall in rigor. By formally melding the "black box" and the "fundamental analysis" approaches, our models offer significant advantages over both.

While many formidable intellects have attempted to formally bridge the gap between the two spheres, until now this has not been computationally possible. The ability to seriously attempt precision data-finding for probabilistic modeling of real-world events, is probably unique to NLP/Black at present, from a technological point of view. Nowhere else in the world is there a well-tested, mature, very-broad-coverage English parser that provides a complete analysis of text, not just syntactically, but at the level of meaning as well.

Using our technology, we can build models that forecast trends within your firm's particular domains of interest. For instance,considerable work has already been done for factors relevant to the mining industry. Forecasting labor action, predicting country risk on a more broad-based basis than is traditionally done, and predicting fiscal and political crises, are other examples of functions our models can provide. Alternately, we can provide the raw data streams (bit vectors for identified textual events) for incorporating into your own formal model.

We will be happy to discuss applications of this technology to meet your data and modeling needs.

Contact: Ezra Black (ezra.black@epimenides.com, 1-718-549-2714)