Part 1 of a series on how Blackworks Capital is building and testing an AI layer for its systematic framework, and what we are learning while we build it.
A note on what this is, and what it is not. This is the first piece in a series, and the series is a working journal rather than the announcement of a finished product or a new capability. The AI layer I am going to describe does not trade live fund capital. It runs in a paper environment, in what I would honestly call beta, while we gather the forward data we need to prove the design out or to disprove it. I am writing about it now, in the middle of the build, because the most useful conversations I have are with other people working on the same problem, and because how a manager approaches a technology like this tells you as much as the technology itself.
With that said, I will state the conviction the series is built on, because I would rather put it on the table than soften it into something no one could disagree with. I believe artificial intelligence will become the dominant method for managing systematic investing.
The reason is narrower than the headlines, and I think it is more durable for being narrow. It has nothing to do with today’s rules-based methods being insufficient. They are not. The deterministic strategies we run manage real capital, including a meaningful portion of my own, and they do the job they were designed to do. It also has nothing to do with AI producing dramatically better returns. That is a marketing claim, and I have no evidence for it at this point. The case rests on something more structural. Markets produce an enormous volume of information every day, and a growing share of it arrives as language rather than as a clean number: central bank statements, earnings calls, geopolitical headlines, regulatory filings, the way a policymaker phrases a single sentence. A machine can take in far more of that material, from many more sources, closer to real time, than any desk of people can with lower costs, and it can read it into a fixed framework with the same unemotional discipline that systematic investing requires in the first place. There are too many structural reasons pulling in that direction for it not to become the dominant approach over time. Whether that takes one year, three, or five, I do not know. The timeline is uncertain. The direction is not.
Stating that conviction changes nothing about how we proceed. We are moving slowly, with the same rigor we apply to everything already in the live book. Believing the direction is real raises the bar on doing it carefully, because the temptation to rush is strongest when you think you can already see where you are going.
I have to draw one line clearly before I go further, because the claim I am making, that AI will come to run systematic investing, invites the exact opposite of what I mean. When people hear it, many of them picture someone pointing a chatbot at the market each morning and asking it what to buy. That is not what we are doing, and I want to say plainly that it would not be systematic investing even if it worked.
Systematic investing is a defined, repeatable process. The same inputs produce the same decision today and a year from now, because the decision comes from a written-down method rather than from whoever is in the chair and whatever mood they are in. It can be tested against decades of history, or parts of it can, more on that later, because it is fixed enough to run on the past. And when it is wrong, you can find out why, because every step is on the record. The discipline is the whole point. A process that cannot be repeated, tested, or audited is not a system, no matter how sophisticated the thing producing its answers.
By that standard, asking a model what to buy this morning is about as systematic as asking your neighbor. The model might hand you a better answer than your neighbor would. It is still a one-off opinion, unrepeatable and untestable, in a more confident voice. Using AI does not make a process systematic. The burden runs the other way: the process has to make the AI systematic. So we do not ask the model for an answer and act on it. We hold it to the same standard we hold every rule we already run, a fixed framework, defined inputs, and an output structured enough to compare from one day to the next and to pull apart when it fails. Everything in this series is about how we impose that discipline on something whose natural instinct is to give a fluent answer and move on.
The first thing to understand about how we use AI is that the framework came first. The Five Forces is not something we reverse-engineered to give a language model something to do. We arrived at it years ago, before AI was a serious thought for us, because it is how we believe markets actually work, and we have written about it at length in our Five Forces material. F1 is macro, F2 is technical and price, F3 is options, positioning, and derivatives, F4 is fundamentals, and F5 is sentiment and human psychology. Because the framework is our view of the world rather than an accessory to the technology, any use of AI has to earn its place inside it. The framework does not bend to fit the model; the model has to fit the framework, or it does not get used.
It also settles what the AI is for. Even before any of this, our deterministic strategies already read all five of the Forces. They read the macroeconomic Force through the relationships between assets, how bonds are priced against equities, how credit behaves, how one sector trades against another. They read sentiment through extremes in price and volatility. These are not narrow instruments.
What the form of a deterministic rule imposes is not blindness. It is distillation. To be fast and testable across decades of history, and to stay free of the operator’s mood on any given morning, a rule has to compress each Force into a clean quantitative signal. That compression is what makes the rule trustworthy, and I would not give it up. Distillation carries a cost that is easy to miss, though: the rule reads that risk is rising without holding the context of why. It can register that credit spreads are widening and that equities are unsteady, and act on it correctly, without ever knowing whether the cause is a banking scare, an inflation surprise, or a single geopolitical shock the market will digest in a week. The number is right. The reason behind the number has been compressed out, by design.
This is where the right tool for the right job becomes the governing idea, and it is the whole argument. A deterministic rule is the correct instrument for distilling a Force into a precise quantitative signal. Nothing about AI changes that, and the AI layer does not try to. An AI is the correct instrument for a different task it happens to be very good at, which is multi-context language reasoning. By that I mean taking in many data points that arrive as language, that are sensitive to context and do not reduce cleanly to a single value, holding that context rather than discarding it, and reading those inputs into the same Five Forces framework to produce additional signals a rule was never the right tool to produce.
I want to be explicit about what the AI is not, because the distinction matters more than any feature I could list. It is not a set of smarter rules, and it is not a replacement for the deterministic strategies. The live book remains exactly as it is. The AI layer is a second, independent reading of the same world, built to do a job the first instrument was never designed for.
Because the two instruments read the same Forces through different inputs and different methods, the signals they produce are orthogonal to one another. That orthogonality is the part I find most interesting as a portfolio manager, because it is the raw material of real diversification. Two views that agree because they look at the same input the same way add very little. Two views that read the same underlying reality through independent channels, and sometimes disagree, are worth far more. The disagreements are not a defect to smooth away. They are information.
The same framework that holds our worldview is also what makes the machine usable at all. Point a capable language model at the market with an open mandate and you get fluent, plausible, untethered narrative. It will tell you a story, and the story may even be a good one, but it will not be bounded, it will not be comparable from one day to the next, and it will be nearly impossible to audit, because there is no fixed thing it was asked to reason about. Point that same model at a specific Force, with a specific job and a specific set of inputs, and its reasoning becomes something you can work with. It is held to a known surface. Its output today can be compared against its output yesterday, because both were answers to the same defined question. When it is wrong, you can find out where, and why.
I have saved the hardest part about building with AI for last, because it is what separates honest work from marketing, and it is the reason this series is a working journal rather than a product announcement.
You can backtest a deterministic strategy honestly, and we do, extensively. You hold out data the model has never seen, you walk it forward through history, you measure whether the edge survives contact with periods it was never fit to, and you accept the verdict. The quantitative pieces of what we are building can be tested that way, and they are.
The reasoning layer cannot be. There is no honest way to reconstruct how the framework would have read the news a year ago, or five years ago, because there is no way to put a language model back into the past without the contamination of hindsight. Ask a model today what it would have concluded from information as it arrived in some prior moment, and it already knows, at some level, how the story ended. The knowledge of the outcome leaks into the reasoning. What it produces is not a faithful reconstruction of a real-time read. It is a memory dressed up as a forecast, and I will not validate anything important on a memory dressed up as a forecast.
That leaves one honest path: forward paper-testing, in real time, watching what the system does as events it has never encountered actually arrive. This makes validating the reasoning layer slower and structurally different from building a deterministic strategy, where you can compress years of history into an afternoon of computation. Here the proving ground has to be built one real day at a time. That is a real limitation that can either be accepted, or ignored, but when the capital at risk is, first and foremost, my own, acceptance of that limitation sets the standard for how patient I am willing to be.
This first piece is the philosophy and the flag in the ground. The rest of the series gets specific about how we are building the AI system and how we keep the system honest while we do.
I will write about the data the model is allowed to think about, and why deciding what to exclude matters as much as deciding what to include. I will write about why a language model cannot be made to vote the way our rules do, and how the reading it produces has to become a structured view of its own. I will write about the problem of making a fluent model trustworthy, since fluency and reliability are not the same thing, and the gap between them is where most of the work lives. I will write about the adversary we have built into the system, whose only job is to argue that we are wrong before any signal is taken seriously, and about how the AI’s role is allowed to grow over time, and the standards it has to clear before it grows at all. Then the series gets specific about the destination. The honest version of my conviction is not that AI improves every strategy equally; certain kinds of strategies are built to use what this technology does best, others gain very little, and the difference tells you where a machine should be pointed first. The later pieces are about that answer, including the strategy we are building around it.
I want to end where I began, because the two halves of my view belong together. I believe AI will become the dominant method for managing systematic investing. I also believe, as the person responsible for managing risk in this fund, that there is a great deal of risk in how AI gets implemented inside a systematic strategy, and that much of that risk shows up when a manager is in a hurry to claim the future before the work supports it. So for now, and for the foreseeable future, this is where AI fits at Blackworks Capital: we are building it, we are testing it, and we are evolving it. It earns a role in managing money only if, and when, it can meet the same rigor, discipline, auditability, and standards we already demand of every strategy we run. We think AI will become the dominant method. For us, today, it is not, and I am not going to pretend otherwise. I am, at the same time, very excited about the future and about continuing to build with this incredibly powerful new technology.
I hold all of this loosely enough to be disproved and firmly enough to keep building. If you are working in this direction too, I would welcome the conversation.
Blackworks Capital LLC manages funds through Blackworks Capital Management LLC, an Exempt Reporting Adviser. Nothing here is an offer or solicitation or investment advice. The systems described are in research and paper-testing and do not manage client capital. Past performance does not guarantee future results.