Teaching P34 to learn markets, not memorize them

Most machine-learning systems are trained to answer a relatively simple question:

Given examples from the past, what outcome should I predict next?

For P34, that is not enough.

The markets we care about are partially observed. A business typically knows what happened to the deals it executed, but it does not know what would have happened to the thousands of opportunities it rejected. Its historical data is therefore biased by its own previous decisions.

That creates a difficult problem. A sufficiently powerful model can become very good at reproducing the patterns present in historical data while becoming dangerously wrong about the opportunities it will encounter after deployment.

Our training objective is therefore different. We are trying to teach P34 to learn how to become right about a market, rather than simply teaching it a collection of historically correct answers.

We describe this informally as learning to learn.

This changes what a training example means.

At the lowest level, P34 can still learn ordinary relationships between telemetry and economic outcomes. A deal has characteristics; some outcome such as profit is eventually observed; the model can learn from that relationship.

But P34 adds higher levels of supervision. The model can make a prediction or construct a candidate portfolio. We can then expose what actually happened in a withheld future period or synthetic continuation.

Was the model right about its own prediction?

And there are several ways to define "right."

• Did the selected portfolio generate the profit the model expected?
• Did supposedly profitable trades actually become false positives?
• Did a proposed interpretation remain calibrated in a different market regime?
• Did the model correctly decide to avoid a market?
• Did it choose the right quantity?
• Did the portfolio remain economically useful when the historical business policy was no longer filtering opportunities for it?

Training can therefore operate on more than the original economic labels. It can use holdouts, forward windows, out-of-fold predictions, portfolio calibration, false-positive outcomes, and other augmented signals describing how the learner behaved.

The object being learned gradually moves from:

toward something closer to:

This is the sense in which P34 is being trained to learn to learn.

P34's architecture is intentionally structured to resist the assumption that the historical dataset describes one complete, observable truth. It can instead construct multiple plausible interpretations — or universes — of the partially observed market.

Different universes can represent different explanations for:

• missing opportunities;
• historical selection bias;
• market regimes;
• false-positive risk;
• changing margins;
• hidden demand;
• and future drift.

Those interpretations can then compete. The training problem becomes partly a problem of learning:

Given what happened when similar interpretations were used elsewhere, which interpretation should we trust now?

That is deliberately different from fitting one large predictor to one historical table.

The higher-level selector works with generalized telemetry about model behavior — regime information, prediction quality, calibration behavior, false-positive correction, portfolio characteristics and related signals — rather than depending entirely on the identity of a particular product or market.

The architectural objective is important: information about how to learn successfully should be more reusable than information about one particular training market.

We also distinguish learning a market from choosing how aggressively to operate in it. Those are separate questions.

First, we want the model to correctly understand the economic decision problem:

• which opportunities are likely to make money;
• which are likely false positives;
• how quantity changes expected outcomes;
• when the market is different from its historical context;
• and when declining to trade is the best action.

Then we can define preferences over economically valid strategies. For example, two users might use the same underlying model but have different preferences for:

• available capital;
• acceptable maximum drawdown;
• portfolio concentration;
• number of losing trades;
• downside variance;
• turnover;
• or the tradeoff between total profit and consistency.

A strategy generating $1 million while suffering a $900,000 drawdown is not equivalent to a strategy generating $800,000 with a $100,000 drawdown. Similarly, two strategies with the same total profit can behave very differently if one earns it through hundreds of highly volatile losses and gains while another produces much more stable portfolios.

Our alignment process therefore looks not only at how much total profit was produced, but also at characteristics such as how many trades lost, how losses were distributed, and what capital was required to survive the strategy. These signals let us align the model's portfolio policy to practical capital and risk constraints.

But those preferences should not be confused with market competence. Maximum drawdown is a preference. Capital availability is a constraint. Risk tolerance is a preference.

Knowing how to reason about the market in the first place is a learned capability.

We treat the latter as dedicated model-training work.

The current iteration of P34 includes explicit architectural choices intended to encourage this behavior. The system separates several levels of reasoning rather than collapsing the entire task into one predictor.

This hierarchy creates places where generalization can be tested independently.

• Was the underlying label estimate wrong?
• Was the regime interpreted incorrectly?
• Was a false positive missed?
• Was the wrong universe selected?
• Was the quantity boundary wrong?
• Was a portfolio economically correct but inconsistent with the chosen capital policy?

Those are different errors, and treating them as different learning surfaces makes it harder for the system to hide all of its reasoning inside one highly fitted function. The architecture is therefore deliberately biased toward reusable learning behavior rather than memorization of one dataset.

It does not make overfitting impossible. No architecture can guarantee that. Instead, it lets us construct training and evaluation loops in which overfitting becomes observable — and in which generalized decision behavior can itself become a training target.

There is a useful distinction between a model that has learned many markets and a model that has learned how to learn a market.

The first model may contain an enormous amount of historical knowledge. The second needs to answer a more useful sequence of questions:

• What evidence is available here?
• How was that evidence selected?
• What is missing?
• Which interpretation of the missing data is plausible?
• Which of my predictions should I distrust?
• What changed relative to the environments where this reasoning worked?
• What portfolio follows from this interpretation?
• And given the operator's capital and risk preferences, should I trade at all?

That is the direction of P34 training.

We do not consider it solved. The current model has a limited regime library, synthetic evidence does not prove live-market performance, and the boundaries of cross-market generalization remain an active research question.

But the architectural objective is clear. P34 is not being trained merely to predict profit.

It is being trained to repeatedly discover how to become right about profit from incomplete evidence — and then to translate that learned understanding into an economically aligned decision.