Quant, quantamental and hybrid intelligence

Whitepaper · Quant, quantamental and hybrid intelligence

1. Quant, quantamental and hybrid intelligence#

Classical statistics, deep reinforcement learning and language-model strategies compared with hybrid intelligence in DeOS

Why backtests flatter and what robust means, why deep reinforcement learning runs on a retraining treadmill, nine flaws of language-model strategies, what hybrid intelligence does differently, and why DeOS accelerates quant teams rather than competing.

Summary

2. In brief#

Every investment process turns information into a position. Classical statistics describes one realized path, reinforcement learning retrains on it and forgets, language models sample the consensus about it. Hybrid intelligence in DeOS decides on a causal graph from the firm's own cases, with expert judgment as calibrated evidence and a memory that never forgets, and shows afterwards whether the firm decided well.

01

The comparison framework

3. Four ways to turn information into a position#

Classical quant, the quantamental combination of quant screens and fundamental analysts, machine learning with deep reinforcement learning, and language-model strategies. Each is described by the same six questions, placed on two axes, and then scored against hybrid intelligence in DeOS on eleven dimensions.

Six questions for any approach#

An investment approach can be described without prejudice by six questions. What does it learn from: which data, how much of it, and whose? How does it treat uncertainty: does it produce a point, an interval under an assumption, or a distribution it can be held to? How does it adapt: by re-estimation, by retraining, by reading, or not at all? How does it use the people in the firm: as specifiers, as prompt writers, or as sources of measured evidence? How is it validated: by an in-sample fit, a backtest, a simulator, or a record of decisions judged against the probabilities they were made under? And what does it cost to run, in people, pipelines and calendar? The same six questions are asked of every approach below, including DeOS.

The four approaches, in one paragraph each#

Classical quant is the statistics of returns: factor models in the Fama and French (1993) tradition, time-series and cross-sectional regressions, risk models and optimizers in the Markowitz line. It is parsimonious and interpretable, and it assumes linearity, stationarity and tails that markets do not have. Quantamental is what most active firms actually run: quant screens and risk models supply candidates and exposures, fundamental analysts supply the reading of filings, calls and management, and a portfolio manager reconciles the two in a meeting. Its weakness is not either half but the join: the analyst's judgment is never recorded as a prediction, so it cannot be calibrated, and the model's output is translated into prose, so it cannot be interrogated. Machine learning and deep reinforcement learning promise to learn nonlinear structure from many features, and in asset pricing the promise has evidence (Gu, Kelly and Xiu 2020). The reinforcement-learning version, a policy trained in a simulator of one history, runs into sample inefficiency, simulator overfitting, instability, catastrophic forgetting and opacity, which chapter 04 takes in turn. Language-model strategies ask a pretrained text model for a view, a score or a portfolio. They are the cheapest to build and, as chapter 05 argues, not strategies at all.

Hybrid intelligence with DeOS is a different construction rather than a fifth model. Decision optimization is Strategy plus Intelligence plus Platform: the firm's strategy formalized into decision policies, an intelligence flow of Perceive, Diagnose, Predict and Recommend run by one pretrained in-context learner with the firm's experts as calibrated sources, and a platform of five capabilities (Governance, Context, Analysis, Recommendation, Learning) that runs every decision and learns from it. The quant's models are components inside it; chapter 06 gives the detail.

The scorecard on eleven dimensions#

The two axes compress eleven dimensions, and the scorecard below opens them up. Reading across a row shows how differently the approaches answer the same question; reading down a column shows the pattern of each approach. The classical column is mostly amber: it handles each dimension under an assumption that holds on average and fails in a crisis. The deep reinforcement learning column is mostly red because its answers depend on a simulator of one path. The LLM column is red for a different reason: a frozen text model has no mechanism for uncertainty, causality, regime change or learning from the fund, and the prose hides that. The DeOS column is green by construction; the proof of each is a walk-forward record on the firm's own decisions. Quantamental is not a separate column: on every dimension it inherits the classical column and adds unrecorded judgment, which chapter 08 is about once that judgment is recorded. No column is wrong everywhere. Classical statistics is the right tool for risk and attribution, a well-simulated execution problem is a fair use of reinforcement learning, and a language model is an excellent reader. The scorecard is about the decision: which approach can take a position under uncertainty, in a changing world, with the firm's own judgment and memory, and show afterwards whether it decided well.

02

Backtesting, overfitting and robustness

4. Why backtests flatter, and what robust means#

Every approach in this paper is validated against history at some point, and history is a small, single, non-stationary sample that has been searched by thousands of people. The failures are well documented. The question is what a validation that could be trusted would look like.

The factor zoo and multiple testing#

Harvey, Liu and Zhu (2016) catalogued several hundred factors published as explanations of the cross-section of expected returns and argued that, given how many had been tried, the usual t-statistic threshold of two is far too low; a newly proposed factor should clear something closer to three. The point is statistical, not polemical: when many hypotheses are tested on the same data, the best of them looks significant by construction. Jensen, Kelly and Pedersen (2023) answer that most published factors do replicate once they are treated as a family rather than as independent discoveries, which is the fair reading of the evidence: the factors are real, shared, and smaller than their discovery papers said. Either way, the number of trials belongs next to the result, and in most research processes it is not recorded.

Backtest overfitting#

Bailey, Borwein, López de Prado and Zhu (2014) showed, with the arithmetic of order statistics, that the expected maximum Sharpe ratio across N independent backtests of worthless strategies grows with the logarithm of N, and that a researcher who tries enough variants on a few years of data will find one with an impressive in-sample Sharpe ratio and a negative out-of-sample one. Panel A of the figure draws that arithmetic: after about ten trials the best worthless backtest sits one and a half standard errors above zero, after about a thousand it sits above three, and a t-statistic of two proves nothing about it. Panel B draws what the researcher then experiences: a family of variants, the one with the best in-sample fit kept, and that variant's live record, which is what the other variants looked like all along. The deflated Sharpe ratio (Bailey and López de Prado 2014) corrects for the number of trials, the length of the sample and non-normal returns; the probability of backtest overfitting (Bailey et al. 2017) estimates, by combinatorial cross-validation, how likely it is that the best in-sample configuration underperforms the median out of sample. Both need the one thing most research processes do not keep: every trial. DeOS keeps every analysis variant in the decision memory, so the N is known when the result is read.

Decay, look-ahead, survivorship and regime#

McLean and Pontiff (2016) measured what happens to a predictor after it is published: returns fall substantially, part of it the correction of the original selection effect and part of it arbitrage by investors who read the paper. A signal that is shared is a signal that decays, which is the first reason an edge has to be specific to the firm. Look-ahead bias is the use of information not available at the time, and it has acquired a new form: a model pretrained on text written after the test window has read what happened (Glasserman and Lin 2023), and the leak cannot be removed from the weights afterwards. Survivorship bias tests today's universe on yesterday's dates, without the names that failed. Regime dependence is the quietest of the four: a sample of fifteen years of daily data is long in days and short in regimes, and a strategy tested on one rate cycle and one market leadership has been tested once. Arnott, Harvey and Markowitz (2019) turned these into a protocol for the era of machine learning: state the hypothesis before the test, keep every trial, use point-in-time data, hold out by time, and treat an untheorized result as a hint.

What robust means: the ladder#

Robust is a ladder with five rungs, and none of them is a higher Sharpe ratio. An in-sample fit proves nothing about the future and hides the search that produced it. An out-of-sample test on one held-out period proves that the strategy survived one draw, and still admits look-ahead through the data and, for pretrained models, through the weights. Walk-forward, point-in-time evaluation re-decides at every date with only what was known then, which closes the leak but says nothing about whether the model's stated uncertainty means anything. Calibration of the predictive distribution does: of the outcomes the model put in its lowest decile, did about a tenth fall there? A point Sharpe ratio is one number from one path; a predictive distribution can be checked every period, per analysis, per person and per model, with Brier scores and reliability curves, and calibration can be measured within months, long before a track record is statistically meaningful. Decision quality is the top rung: calibration, policy adherence and conviction against outcome, recorded at decision time so hindsight cannot rewrite them, which separates a good decision from a lucky one. A backtest stops at the second rung. DeOS evaluates at the top two, with every trial kept and the proof plan written before the first decision runs: the plan names the KPI, the counterfactual (the old process on the same cases) and the gate.

03

Classical statistical models

5. What they do well, where they break, where they stay#

Factor models, regressions, time-series and cross-sectional statistics are the foundation of quantitative investing and of every risk report an investment committee reads. They deserve a fair hearing: they are honest tools that state their assumptions, and the assumptions are the problem.

What they do well#

A linear factor model has a handful of parameters, each with an economic meaning and a standard error. It is hard to overfit by accident, easy to explain to a board and a regulator, and it comes with a complete theory of inference. For the questions it was built for, what exposures does this book carry, what did they earn, how much was the market and how much the manager, nothing has replaced it. A coefficient is a sentence: this exposure earns this premium with this standard error. The sentence is true under the assumptions that produced it, and the figure below is about those assumptions.

Where they break#

The assumptions are the limits, and the three panels are three of them. Linearity (panel A) misses thresholds, saturation and interactions, and a crisis is where the nonlinearity lives; the point below the band is not noise, it is the event the model has no term for. Normal residuals (panel B) call the moves that matter impossible; Mandelbrot (1963) made the point about cotton prices sixty years ago, and a single day in October 1987 made it about equities. Variance is a single, intuitive number, and it is the wrong number for tail risk. Stationarity and the single path (panel C) assume that the process that generated the sample is the one that generates the future; Lo (2004) describes markets as adaptive, with relations that appear, are learned and fade. The deepest

limit is not statistical: the model describes one realized path and has no way to ask what would have happened, and it forecasts a return without modeling what the firm's own action does to the price, the position or the outcome. A regression is a prediction tool; a decision needs a model of actions and outcomes. Two further limits do not draw well: hand-crafted factors are shared, so their premia are shared and decay once traded, and finding out that a new specification is needed takes a cycle, because the re-specification is by hand and starts after the break has become visible.

The quantamental join#

Most active firms run classical statistics next to fundamental analysts and let a portfolio manager combine the two: the quantamental approach. The combination is sensible and the join is where the value leaks. The analyst's view enters as prose and a conviction word, so it is never calibrated; the model's output enters as a number, so it is never explained; the two meet in a meeting whose record is a deck, so the decision cannot be replayed; and when the outcome arrives, nobody can say which half was right. The join is a communication gap, and the fix is a shared object that both can read and both can be measured against, which chapter 08 proposes.

Where they remain the right tool#

In DeOS, classical models are components. The risk model, the factor exposures, the attribution and the benchmark neutralization plug in as analysis archetypes behind a trust gate, with a calibration record like every other input, and their outputs become drivers on the causal decision graph. The firm keeps what it has built and can explain, and stops asking a regression to decide. The test for any model, classical or not, is whether it can say how wrong it might be, in the units of the objective, about the situation in front of it: a standard error under normality is an answer under an assumption; a calibrated predictive distribution, checked against outcomes by bucket, is an answer under evidence. Every classical tool has a place in one of the four analysis components, keeps its mathematics, and gains a calibration record, a trust gate and a position on the graph.

04

Deep learning and deep reinforcement learning

6. The promise, the treadmill and the alternative#

Machine learning learns structure that hand-crafted factors miss, and in asset pricing the evidence for that is real. Deep reinforcement learning promises more: a policy that learns to act. In finance it meets a world that is non-stationary, happened once and charges for every experiment. In-context learning is the alternative, and it changes the question.

The promise#

Gu, Kelly and Xiu (2020) compared regression trees, random forests and neural networks with linear models across a large set of predictors and found economically meaningful gains in out-of-sample return prediction, concentrated in the nonlinear interactions that linear models cannot represent. Nagel (2021) sets out why this is both real and limited: the gains are modest in absolute terms, the signal-to-noise ratio of returns is low, and the models are only as stable as the sample they learned. The lesson is that flexibility helps when it is disciplined, and that the discipline has to come from outside the fit.

The practical failure modes of deep reinforcement learning for trading#

Reinforcement learning learns a policy, a mapping from state to action, by interacting with an environment and maximizing a reward (Sutton and Barto 2018). Applied to trading, the environment is a simulator built from one historical path and the reward is a return net of estimated costs. Each step of that sentence contains a failure mode, and they compound. Sample inefficiency is the root: a deep policy needs millions of interactions, a market offers a few thousand days of one history, and the same path is replayed until the policy has memorized its accidents (Sutton and Barto 2018; Henderson et al. 2018). Reward hacking follows: the policy maximizes the reward as implemented, including every simplification in the simulator, fills, impact, borrow and latency, so backtested returns come from the simulator's holes (Amodei et al. 2016; Zhang, Zohren and Roberts 2020). Instability: Henderson et al. (2018) showed that results of the standard algorithms vary materially with the random seed, the hyperparameters and the code base; two identical research runs give two strategies, and a backtest of a DRL strategy is one draw from that distribution. Catastrophic forgetting is the oldest problem (French 1999): training on new data overwrites what the old data taught, and the remedies (Kirkpatrick et al. 2017; Hadsell et al. 2020) are partial. The black box means the committee cannot interrogate a position and the PM cannot override a reason they cannot see (Lipton 2018). Exploration costs money live, so it is done in the simulator, which brings the sim-to-real gap back: live performance diverges exactly where the simulator was wrong (Zhang, Zohren and Roberts 2020).

The retraining treadmill and catastrophic forgetting#

In a market the failure modes combine into the retraining treadmill: trained on regime A, the policy fails when B arrives; retraining needs B data that does not exist yet, and the gap is paid in live capital; when the retraining finally runs, A is overwritten and must be learned again when it returns. The figure draws the skill of one model on two regimes over time. For the retrained model, skill on A collapses during the retraining on B and skill on B collapses when A returns, so the model is always one regime behind and never holds both. The in-context learner keeps its skill on A because nothing was overwritten, builds its skill on B from a few cases while its stated uncertainty widens and trust is withheld, and recalls A by similarity when A returns. Each retraining run also gives a different policy, because of seed sensitivity; the in-context learner is deterministic given its context and memory, so the same question gets the same distribution.

Contrast: in-context learning#

In-context learning is the ability of a pretrained model to solve a new task from examples placed in its input, with no change to its weights (Brown et al. 2020). Prior-data fitted networks make this rigorous for prediction: Müller et al. (2022) showed that a transformer trained on synthetic tasks drawn from a prior learns to approximate the Bayesian posterior predictive for a new task given its examples, in one forward pass; Hollmann et al. (2025) showed in Nature that such a tabular foundation model, TabPFN, predicts on small datasets at or above the level of tuned gradient-boosting ensembles, in seconds, with the full predictive distribution as output. Nothing in the market under test is in its weights: the prior comes from synthetic and cross-domain tasks, and the firm's own cases are the context. Against the retraining loop, the differences are the ones a decision needs: when new data arrives the context changes and the weights do not; nothing is forgotten; a handful of cases is enough, because the learner was pretrained to infer from small samples; adaptation is one forward pass rather than a training run, a validation and a deployment; the output is quantiles, tails and spread rather than a point or an action; synthetic counterfactual periods can be given as context, which a training set that is the realized path cannot represent; and the learner is deterministic given the same context and memory, where a policy varies by seed, run and data order.

Two consequences follow for the decision problem. Small data is enough. A firm with forty comparable decisions in its memory has forty examples, which is a rich context for a learner built for small data and nothing for one that needs millions of transitions; the prior answers the zero-shot case with a wide, honest distribution. Trust is earned, not assumed. DeOSAlpha, the Decision ICL Engine, places trust gates and calibration records in front of every component: nothing learned decides until it has shown skill on a walk-forward record, a failed gate hands control back to a simpler decider, and novelty against the memory pushes the recommendation to wait.

The world that didn't happen#

History happened once, and every model in this chapter learned its luck as structure. DeOS treats the paths not taken as a first-class input: counterfactual simulation on the causal decision graph answers what would have happened under a different action, and synthetic counterfactual periods, generated by a world model and filtered for plausibility, are given to the learner as context so that the search and the stress tests cover regimes the sample never showed. They are context, never evidence: only realized outcomes enter the calibration record. The firm's few real cases, plus the many worlds that did not happen, are the training context of every analysis: the difference between a model that has seen one crisis and a decision system that has rehearsed many.

Where reinforcement learning still fits#

05

Language-model strategies

7. Language-model strategies: the craziness and the flaws#

A language model is the best reader ever built and a poor investor, and the industry is confusing the two. This chapter takes the flaws one at a time, with the theoretical reason and the evidence for each, says why an LLM strategy nevertheless seems to work right now, and then gives language models the role they are good at inside a decision system.

The argument is not that language models are useless in investing; chapter 06 and chapter 08 use them. The argument is that a strategy whose engine is a language model, whether it scores sentiment, picks stocks, sizes positions or plays the role of a famous investor, fails on nine counts that are properties of the architecture, not of the current generation of models, and that the apparent success of such strategies in the current market is the signature of the failure, not evidence against it. The figure below places the flaws on the chain that every LLM strategy runs, from the prompt to the return; the sections that follow take them in order.

Why it seems to work now#

Three conditions have held at once: a long bull market, leadership concentrated in a handful of large technology names that are the most written-about companies in the corpus, and momentum rewarded. Under those conditions the consensus view is the winning view and a model that returns the consensus looks like skill. None of the three is a property of the model; all three are properties of the period, and periods end. The honest test of an LLM strategy is its performance by regime, its exposure to beta and crowding, and whether its view changes when the facts change rather than when the prompt does.

The right role of language models#

A language model reads faster and more widely than any analyst, extracts entities, events and stated numbers from documents, summarizes and turns a chain of reasons into prose. Inside DeOS those are its jobs, each with a human / AI mode and a guard, and none is the engine: the recommendation, the probability, the conviction score and the why chain are computed on the causal decision graph by the Decision ICL Engine, and the language model explains the computed why chain rather than inventing one. A strategy in which it does more is not a strategy but a sample from the consensus with a confident voice.

Six tests for any process that uses a language model#

Each test takes an afternoon, needs no data science, and exposes one of the flaws above. A process that passes all six is using the model as a reader. A process that fails any of them is using it as the engine.

06

Hybrid intelligence with DeOS

8. What hybrid intelligence with DeOS does differently#

Decision optimization is Strategy plus Intelligence plus Platform. The intelligence is hybrid: one pretrained in-context learner and the firm's experts, each as a calibrated source, running Perceive, Diagnose, Predict and Recommend on a causal decision graph. This chapter states what that construction does that the four approaches cannot, property by property.

Probabilistic by construction#

Every predicted quantity is a distribution, never a number: quantiles, tails and spread for every candidate action, in one forward pass of the learner. The distribution is decomposed, and each part goes to a different place in the decision, which is the routing rule of The world that didn't happen: hedge chance, investigate ignorance, step back from novelty. The spread of the posterior members' means is the epistemic part, what more data or an expert would change; the mean of their spreads is the aleatoric part, what the world does on its own (Hüllermeier and Waegeman 2021); the distance of the present context from the decision memory stands in for the unknowable. Aleatoric uncertainty goes to sizing: position size, hedges and CVaR on the predicted quantiles. Epistemic uncertainty goes to learning: the value of information, the next question to ask, the expert to elicit. Unknowable uncertainty goes to waiting: the stated uncertainty widens, trust is withheld, control hands back to a simpler decider, and wait is a valid recommendation. A standard error under normality reports one number for all three and sends it nowhere in particular.

Causal#

A decision is an intervention, not a classification (Pratt 2019). DeOS represents each decision on a causal decision graph: the actions the firm can take, the drivers they move, the outcomes that follow, the goals and constraints they are judged against. Causal diagnosis finds where the firm stands on the graph; counterfactual prediction computes what would happen under each action, on the graph, which is Pearl's second and third rung (Pearl 2009) rather than a description of them; contextual optimization chooses within the mandate. A regression predicts the outcome variable; the graph predicts the outcome of the action, which is the quantity a decision needs.

Small data and the world that didn't happen#

The learner is pretrained on synthetic and cross-domain tasks, not on the market under test, and infers from the firm's own cases placed in its context: the past decisions most similar to the present one, selected by the goals and the perceived context, with their outcomes. Forty comparable decisions are a rich context for a learner built for small data. Synthetic counterfactual periods, generated on the causal decision graph and filtered for plausibility, extend the context to regimes the sample never showed; they steer search and stress tests and are never counted as evidence.

Human judgment as structured, calibrated evidence#

Experts are not replaced and they are not prompts. Their judgment is elicited as structured predictions on precise questions, a range and a confidence, recorded at decision time so that hindsight cannot rewrite it, scored into a calibration record per contributor (Tetlock and Gardner 2015), and fused with the models by debiased weight that rises where the expert has been right and the data is thin. Noise between experts and across occasions (Kahneman, Sibony and Sunstein 2021) becomes measurable and is reduced by structure. The mode is set per component: Automated where data is rich and errors cheap, Augmented where data and judgment both carry information, Judgement only where accountability, values or novelty require a person. The decision itself always stays with the named decision owner.

Decision memory and continuous learning without retraining#

The decision memory holds every decision with its context, its analyses and their variants, the probabilities and the conviction score at the time, every override with its reason, the outcome, and the decisions not taken. It is the context of every future analysis and the record every component and person is calibrated against. Learning is the update of that context and those weights, not a training run: the standardized error of a new outcome under the current fit is the engine's surprise; a sequential test on it raises the drift alarm, resets the trust region and drops the rows before the change point. No weights change, nothing is forgotten, and the old regime is recalled by similarity when it returns.

Governance, audit and evaluation by decision quality#

Every recommendation is replayable: which data as of which date, which analysis in which version, which component in which mode, which person with which prediction, which override with which reason; the Governance capability keeps that trail across solution, decision policy and analysis. Evaluation is by decision quality: calibration, policy adherence and conviction against outcome, judged against the probabilities the decision was made under rather than the outcome alone. It separates skill from luck, is visible within months at the analysis level, and is what an allocator can read instead of a track record.

What the other approaches become inside DeOS#

None of the approaches in this paper is discarded. Each becomes a component with a defined role, a human / AI mode and a trust gate, and none of them is the engine.

07

Synthetic data from the world model

9. The data that does not exist, and how it is made safe to use#

Every approach in chapters 02 to 05 is limited by the same fact: most of the data a decision needs was never produced. History happened once, the state that matters is hidden, the quantities that would settle the question were never measured, and the data that would help sits in other firms. DeOS generates what is missing with a world model and lets none of it through without passing five quality gates.

Four kinds of missing data#

The missing data is not one problem but four, and they need different things from a generator. The world that didn't happen is the one realized path: the crises and rotations that nearly occurred left no observations, and chapter 02 showed what a model learns when it treats that path as the population. A partially observable state is the second gap: the regime, the liquidity, the positioning of other investors and the intent of a management team are latent, and what the firm observes is a noisy projection of them; a decision made on the observations alone is a decision made on the shadow, which is the formal setting of Kaelbling, Littman and Cassandra (1998). Data that is not measured is the third: nobody recorded what the deal not done would have earned, what every other candidate would have returned on the same dates, or the conviction the analyst held before the meeting, so the record contains the action taken and none of its alternatives. Proprietary and hidden data is the fourth: deal flow, private marks and order books are in other hands and cannot be bought, and the firm's own record, the one source that is specific to its edge, is small and must not leave its instance.

What the world model generates#

The world model of the Decision ICL Engine is a bootstrap ensemble that learns the dynamics of the decision from the firm's record and the causal decision graph, under one contract: given the context, the action and the goal, what outcome follows and what does the context become. The spread of the members' means is what the model does not know; the residual spread is what the world does on its own, which is the same separation of ignorance and chance that chapter 06 routes to learning and to sizing. Once the ensemble has agreed with reality it is rolled out, in the sense of Ha and Schmidhuber (2018) and Hafner et al. (2020), to produce the four products in the figure: synthetic counterfactual periods that recombine regimes, stress one driver or bootstrap the blocks of the window; latent state samples, several hidden worlds that could all have produced the observations, each carried to its own outcome; hindsight rows and shadow decisions that fill the unmeasured cells with a value and an uncertainty; and augmented cases, plausible neighbours of the firm's own decisions that stay on its instance. The learner is at home with this material: a prior-data fitted network was pretrained on synthetic tasks in the first place (Müller et al. 2022), so synthetic rows in its context are the kind of data it learned to learn from.

The reason for the gates is in the same literature. A world model's errors compound along a rollout, and a policy that is optimized against its own model learns the model's mistakes: Janner et al. (2019) bound the damage by keeping rollouts short, and Chua et al. (2018) keep the ensemble's disagreement in view for the same reason. The simulator overfitting of chapter 04 is this failure in another costume. A learner fed its own dreams without a gate would be a deep reinforcement learning agent with better manners.

Five quality gates#

A generated row has to pass five gates in order before it reaches the learner's context, and the gates test different things. The first is statistical fidelity: the marginal distribution of each variable, its tails and the relations between variables have to match the real data, and tails are where generators fail most, because every method that minimizes an average error regresses toward the middle; a synthetic equity drawdown that arrives without the credit spread widening that goes with it fails here. The second is novelty: a row that is a copy of a real row with noise carries no information and leaks the record, and a row far from anything that could happen is a fantasy, so the row has to sit in a band of distance from the real data, which is the authenticity criterion of Alaa et al. (2022). The third is learnability, which is the causal test: a model fitted on the synthetic rows has to predict real data about as well as one fitted on the real rows (train on synthetic, test on real, in Esteban et al. 2017 and Jordon et al. 2022), and an intervention in the synthetic world has to produce the effect the causal decision graph and the real interventional data show, so that the generator is checked for mechanism and not only for correlation. The fourth is usefulness, and it is the one that matters to the firm: with and without the synthetic context, walk-forward on the firm's own held-out decisions, does calibration improve, does policy adherence improve, does regret fall; a context that only makes the learner more confident is rejected. The fifth is provenance and separation: every row carries its generator, its seed, its gate scores and the real data it was conditioned on, and it never enters the evidence record, a calibration record or a performance claim.

What this gives each approach, and what it does not#

Classical models gain stress distributions beyond the sample, which is what their thin tails lacked, and keep their interpretability because the synthetic periods are labelled and the model is re-estimated on real data only. Deep reinforcement learning gains the thing it always needed and never had: a simulator that is gated rather than trusted, with the ensemble's disagreement and the gate scores deciding how far a rollout may be

believed. A language-model strategy gains nothing, because its material is text and none of the four gaps is a text gap. The quant team gains a data factory with a quality department, and the committee gains a sentence it can hold the system to: synthetic data in DeOS is context, never evidence; it never raises a Sharpe ratio, and its value is the uplift in decision quality that the fourth gate measures on the firm's own decisions.

08

For quant teams

10. For quant teams: an accelerator, not a competitor#

A quant team reading this paper may hear a competitor. It is not one. DeOS is the decision layer the team's models have never had: it takes them to market faster, removes the maintenance that eats the research budget, keeps the firm's uniqueness on its own instance, and puts the quant, the fundamental analyst, the PM and the investment committee in front of the same object.

Faster to market, less to maintain#

A model build runs in months: a data pipeline, features, a model, a backtest harness, validation, deployment, monitoring, and then the retraining that never ends. A decision policy in DeOS starts from an archetype: one of four policy types, the analysis archetypes it needs, the firm's context connected through the Context capability, the quant's components plugged in behind trust gates with a calibration record from day one, and a proof plan agreed at the gate. The first live decision under the policy, with its proof running in parallel to the existing process, is weeks away rather than quarters. The quant's work goes into the edge, the driver that is specific to the firm, rather than into plumbing that every firm rebuilds. The largest hidden cost of a quant platform is keeping it alive: retraining schedules, drift reports, point-in-time databases per study, a harness per strategy, and a regime-change fire drill every few years. In DeOS there are no retraining pipelines, because the learner does not retrain; a new signal is a new component rather than a new pipeline; one in-context learner sits behind every analysis, so an improvement to the learner propagates to every policy at once; and the Learning capability does the drift detection, the calibration and the attribution that the team now does by hand. Expert judgment moves from a meeting outside the model to a structured prediction inside the analysis, and the deck translated from the model is replaced by the same decision policy the quant built, read as a why chain and a conviction score.

Uniqueness kept#

The firm's strategy, data, judgment and decision memory live on a dedicated instance. Nothing learned from the firm's decisions leaves it; no other client's decisions enter it. The quant's own models plug in as components and are never shared. What is common across clients is the learner and the archetypes, which carry no information about any market; what is unique is everything that makes the policy the firm's: the causal decision graph, the context, the experts' calibration records and the memory. The uniqueness argument of chapter 02 and chapter 05, that a shared signal decays and a shared model is consensus, is answered by construction.

End-to-end alignment: closing the communication gap#

The quantamental join leaks value because four roles read four different things. The quant reads exposures, t-statistics and a backtest; the analyst reads filings and writes a thesis in prose; the PM reads two memos in two vocabularies; the committee reads a deck and hears a conviction word. In DeOS all four read one decision policy. The quant's model is a driver on the causal decision graph, with the calibration record of every component beside it; the analyst's thesis is a set of drivers on the same graph and their view is a structured prediction with a calibration record; the PM reads one recommendation with a conviction score, ranked options and the why chain from several perspectives, and overrides with a recorded reason; the committee reads what was assumed, what was predicted, how confident, what would change the answer, and, afterwards, whether the decision was good regardless of the outcome. Nobody translates, because there is nothing to translate.

Figure 17 One decision policy read by four roles. Each reads the part they own in the vocabulary they already use, and each is measured against the same record.

What the quant team keeps, what it stops doing, what it gains#

09

The comparison on one page

11. The comparison on one page#

The argument of this paper as one table. Each cell is a verdict; the chapters above are the evidence.

Deep

© vencortex®. Not investment advice.