PRACTICAL A.I.

 

 

 We use A.I.to investigate accepted knowledge, general financial theory, and computer programming. But what about company specifics, what is the optimal level of inventory for a specific operation; what plant and equipment will be necessary to make a particular product? The specific company data has to be in shape. There are four principles that Professor Vasseur of Cal, Berkeley applies: FAIR data must be:

            ·      FINDABLE. It may be all over the place, in spreadsheets and notes.

·      ACCESIBLE. In disparate computer systems.

·      INTERPRATABLE. It may not exist in a common data format.

·      REUSABLE. It can’t get lost.

Otherwise, the best description of the data will be GIGO, garbage in and garbage out. This is probably the general reason why the adoption of A.I. at the corporate, as opposed to the department, level has been slow. It is expensive to get all the company’s data in this form.

As to what is happening in A.I. right now, technology professor Susan Athey of the Stanford Business School says that (in spite of large A.I. expenditures), there has been no change in GDP growth for a long time. To make A.I. work, you have to restructure; that is what the above implies as well.

A.I. seeks to augment or replace human intelligence with superior machine intelligence. But there is a very large difference between them. A Stanford Alumni meeting on 9/25 summed up the difference very well. A wearable button counterposed Techie vs Fuzzy. Techie, by its (usually linear) model assumptions, is very specific. You can design a specific electrical circuit with that linearity assumption. In contrast, the humanities and modern quantum physics is fuzzy. Into that stew fits the social sciences which come down squarely in the middle. Even in economics, as we have illustrated in our 5/23/25 posting, market behavior is straightforward at times; but fuzzy at others when perceived opportunity beckons. For an investor, to act consistently, it is very important to know which is which. As long-term value investors, we are concerned basically with specific asset prices and returns. But the following paper suggests that short-term stock market behavior can be profitably modeled by A.I. We discuss why.

The field of A.I. is very broad. In general, the mathematician Hala Nelson (2023) writes: “…data fuels machine learning algorithms that in turn power many popular AI and/or robotic systems.” ¹ In specific, it is interesting to trace the development of A.I. from deterministic formulas to statistical probabilities - and generative A.I.

Deterministic Formulas

“In the beginning (around seventy years ago) A.I. research attempted to mimic intelligence using rules and logic.” ² It was believed all that was necessary was to feed machines facts and the logical rules to handle them. “In the late 1990s…machine learning infers the rules from the data…” Thus the growth in “…Bayesian networks, evolutionary algorithms, decision trees, random forests, regression…and others.”  ³ This branch of A.I. measures phenomena by defined or derived formulas that have only one answer.

Statistical Probabilities

Modern physics has since discovered that the world at the quantum level is probabilistic. ⁴ This also extends to generative A.I. Whether you wish to generate the next word in a sentence or produce similar documents, using computers you can generate numerical vectors which can be compared with other vectors going in the same direction. Generative intelligence seeks to match probability distributions. “Generative models are built on the assumption that we can only interpret input data correctly if our model has learned the underlying statistical (distribution) of this data.” ⁵ The reason for the interpretation of data as probability is context.                           

Context

Context is the hallmark of modern life and of neural networks. There is a difference in models between explanation and predictability. An explanation that something is key is the province of deterministic formulas and their formal analytic statistics. Short-term predictability, on the other hand, may include many non-key variables in the model, and thus add nuances. Thus it is possible to justify the addition of many parameters to a neural model to achieve the subtlety of context.

A Trading Study

Kelly (of the Yale School of Management), Malamud (of the Ecole Polytechnique Federale de Lausanne), and Zhou (of the Yale School) published a (2021) paper showing that simple models of stock behavior “understate return predictability” compared with “complex models”, where the number of parameters exceeds the number of observations. We are acquainted with standard regression where the models are simple, where given 80% of the data to train the model one tries to predict the remaining 20% of the data by solving a matrix linear equation. That strategy doesn’t work when the model parameters exceed the data. “In this case, the regressor inverse covariance matrix is not defined. However. the pseudo-inverse is defined…Our paper also relates closely to a growing empirical literature that uses machine learning methods to analyze stock returns. The state-of-the-art market return predictions uses high dimensional models with shrinkage and demonstrates robust out-of-sample predictive power.” You get the idea. The paper involves very complicated mathematics and more than 10,000 model parameters.

We cite this study which indicates A.I. is also applied to stock market analysis, but also to make a point. In comparison, long-term stock investment is simple. One only has to be concerned with long-term earnings, and not with many short-term model parameters that have to be estimated from markets that are assumed to be less volatile than they actually are. Perplexity suggests, “For genuinely robust conclusions it is crucial to combine rigorous quantitative analysis with ongoing qualitative assessments and stress testing for structural shifts and rare events.”

__

Some of our readers might be wanting to run their own A.I. models. Now you can with the Ollama interface that offers around 11 different generative A.I. models. Ask how this possible when companies are spending billions of dollars in cloud-based models. The models you will be getting are “aggressively” derated. But nonetheless, the Ollama models apparently work well enough on PCs that contain at least 4 Intel or AMD cores, 16gb of RAM, and at least 50 gb of SSD storage (maybe disk storage will work as well, but we haven’t tried that out). The Ollama models are also transportable to larger cloud models in the Ollama data center. It is also important to note that there are daily usage limits which are set very low to incent users to pay for more use.

 

Footnotes