Implementation Velocity is Key

Portfolio Cat's core aim is to increase implementation velocity when manging cryptocurrency portfolios.

This article highlights our thinking on AI as a vehicle to protecting digital asset wealth on our own portfolios and as to whether we can profit when combining AI, Automated Rules Management and Manual Intervention.

The Importance of Sensibility over Delivery

Many platforms throw out terms like "AI" and "Automated" as a corollary. Those thinking about AI as a means to trade should read this article.

We focus upon exploring options, testing and determining practical outcomes.

Yet - we keep an open mind whilst avoiding hyperbole.

Is AI Driven Trading Possible?

We see all manner of fervour over AI and the ability to automate profitability.

The idea, that one can set up some agents and point it at a trading exchange to make one incredibly rich sounds appealing.

Perhaps some may manage to do this and do incredibly well, but we must firstly understand what AI is and is not.

Understanding Artificial Intelligence

Artificial General Intelligence (AGI) does not currently exist.

Instead, most modern AI systems fall into two broad categories:

• Machine Learning (ML)
• Large Language Models (LLMs)

Breaking this down:

Large Language Models are non-deterministic. Given the same prompt, an LLM may produce different responses on different occasions.

Machine Learning models are generally deterministic once trained and deployed. Given the same inputs, the same model should produce consistent outputs.

Why this matters for trading?

LLMs may open the door to novel trading insights but yet they carry risk - this is why professional investors are not hooking up LLMs to their trading as fast as retail may believe.

Yet - Large Language Models provide inroads to producing output humans would take significant time to produce.

Why this matters for Portfolio Cat?

We want to harness the novelty of LLMs and yet respect the benefits of deterministic machine learning and human driven development.

Understanding what the technology paradigm really is:

• Engineer Driven (Humans)
• LLM Driven (Humans and Machine)
• ML Driven (Human and Statistical)

Platform Development

• Mostly Human driven
• Testing focused
• Input and Output Focused

LLM Integration

• Coding assists
• Insights
• Trade Rule (DSL) assistance

ML Sidecars

• Hard to implement, yet predictable and testable
• Focused and human driven
• Reproduceable consistently

A Real Example - an Automated LLM Agent based Stock Picker

We recently saw a review by a licensed trader on the potential of LLMs to classify instruments and determine their viability for investment. There was a github repository. They referenced an open-source project that used terms like agents, and other modern industry terms.

The software produces a text summary indicating whether a stock is investable at that point in time.

Our hypothesis on using AI to pick stocks, assets and crypto

Ultimately, we must compare the cost of setting up technology versus existing services.

Professional premium newsletters and stock pickers can be anything between $1000-$2500 a year.

It seems as if the cost per ticker could be anything from $1 to $10.

If running our own independent stock picker, we would need a mechanism to incorporate Retrieval-Augmented Generation (RAG).

The hypothesis, therefore, is that we would not want to build a competitor to well established professional investment market research.

Our hypothesis on using AI to pick stocks, assets and crypto

Ultimately, we must compare the cost of setting up technology versus existing services.

Professional premium newsletters and stock pickers can be anything between $1000-$2500 a year.

It seems as if the cost per ticker could be anything from $1 to $10.

If running our own independent stock picker, we would need a mechanism to incorporate Retrieval-Augmented Generation (RAG).

The hypothesis, therefore, is that we would not want to build a competitor to well established professional investment market research.

Our review of their software told us the following:

• Their software was great technically
• The insistence on integrating LLMs made the software expensive to run
• There was no real ability to evaluate the decisions taken - the outputs were text only
• The operational costs were high
• There were better technologies and approaches to make the software more resilient and easier to maintain
• When reviewing the code, we could only imagine more work to make using their solution viable - Low Implementation Velocity

Our Intuition

Those that professionally build these systems will do the following:

• Assess the cost per call
• Determine frequency - how often do we evaluate the market
• Determine the universe
• Evaluate the system deterministically and repeatedly

Most Modern Approaches a modern interpretation of existing approaches

They mention the word "AI" to hook people into the idea of lower costs

AI has high costs per token exceeding what professional analysts perform for their clients

That expert software engineers and analysts will out-perform true implementation

The Smoking Gun (September 2026)

The video put a cost of producing an instrument trading recommendation at between $1 to £4 per report depending upon the LLM model.

Which when we compare LLM insights to professional guidance is far more expensive - especially if running repeatedly during the day.

When LLMs update, the input (prompts) may need to change - often, LLM providers release new models without giving customers the chance to update their software. This creates regression.

LLM providers are providing the exact opposite of reliability of traditional software providers - yet, with the novelty that traditional software providers cannot provide.

Use AI to work on pre-determined information

We are saying that as much information should be pre-determined before hitting LLMs.

Use AI to build your Portfolio Management Technology Infrastructure

Whether we are using agents to build the solution architecture or humans, we are better placed in using LLMs to build the technology to increase throughput.

How we see LLMS fitting into our technology

We see LLMS as being best suited as to taking our Trading DSL (Trading Decision Support Language) to help build rules, interpreting existing rules, making lives easier for our customers and ourselves.

We would like to be in a position where we have an AI managed portfolio within our own risk appetite profile.

LLMs would act off information we provide to generate Trade rules, remove strategies, create new strategies based upon our risk model.

Local LLMs over centralised LLMs and Portability

It is worth building local LLM based technology first, just to test the process. If these are successful enough, your cost base may be low.

This does not exclude switching calls to a centralised LLM provider in the future.