Cadence
How often the system wakes up, runs the full pipeline, and executes a trade. Our tests use daily (13 bars) and weekly (65 bars) cadences. This determines trading frequency and turnover costs.
Research note — July 2026
Portfolio infrastructure / Agentic systems
The agentic trading hype is real. The rigor mostly isn’t.
Most “autonomous” trading systems require continuous human supervision — tweaked prompts, hand-crafted success criteria, fragile self-improvement loops. We built something different: a simulation-driven, mathematically grounded portfolio engine with no human in the loop after deployment.
2.4+Sharpe ratio
MV MPC, daily H=1, corrected
5Rolling OOS windows
All post-training-cutoff data
0Human interventions
After initial deployment
The problem
Agentic trading is stuck in a loop — literally.
The courses and frameworks dominating the agentic trading conversation share a common structure: an LLM agent, a set of manually defined goals, a scoring function someone had to write by hand, and a self-improvement loop that periodically fails in ways nobody anticipated. Every deployment requires a human to define what “success” looks like, monitor whether the agent’s interpretation matches the intent, and intervene when it drifts.
This is not autonomy. It’s a more expensive form of oversight.
The deeper issue is architectural. LLM-based agents generate plausible-sounding decisions, but they cannot guarantee constraint satisfaction, cannot prove they are optimizing the objective you specified, and cannot produce an audit trail that a compliance officer or allocator can trust. When markets move fast, “the LLM said so” is not a risk management framework.
The question we started with: Can you build a trading system that is genuinely autonomous — not just automated — with mathematically enforceable risk constraints, no prompt engineering, and performance that can be verified against real out-of-sample data?
The architecture
Two layers. One pipeline. No human required.
CausaQuant is built on a two-layer architecture. The first layer is a proprietary generative simulator, trained on 2020–2025 intraday data for AAPL, MSFT, NVDA, and GOOGL, that learns the joint return distribution — including realistic co-movement, volatility clustering, and tail behavior — of the chosen asset universe. The second layer is a Model Predictive Control (MPC) optimizer that consumes the simulator’s output and produces optimal portfolio weights at each rebalance point.
No LLM. No prompt. No self-improvement loop. The simulator is frozen weights; the optimizer is a solved mathematical program. Both are deterministic given their inputs.
SelectAsset universe
SimulateGenerative model, frozen weights
SampleK=200 forward trajectories
OptimizeMPC solver, MV or RP
ExecutePortfolio weights
The user’s only decision is which assets to include. After that, the system runs entirely on its own — pulling real market data, generating forward scenarios, solving for optimal weights, and executing the rebalance. No dashboards to monitor. No thresholds to tune. No prompts to revise.
Design decision
Most systems blur two clocks that should be separate.
A critical architectural distinction in CausaQuant is the explicit separation between when the system wakes up to trade and how far ahead it looks when it trades. These are different questions with different optimal answers, and conflating them is one of the most common sources of instability in time-series optimization systems.
Horizon H
How far ahead the simulator projects, and which future checkpoints the MPC optimizer actually uses for its objective. H=1 means the optimizer cares only about the next decision point — ignoring the noise in between.
This separation matters empirically. As you’ll see in the results, the interaction between cadence and horizon is the primary driver of performance stability — and the system’s transparency about this interaction is what makes it auditable.
Results — Horizon sensitivity
Prediction quality degrades with distance. The system knows this.
We ran the full pipeline across six configurations — three daily-cadence horizons (1, 2, 3 days ahead) and three weekly-cadence horizons (1, 2, 3 weeks ahead) — over a fixed out-of-sample window (Jan–Apr 2026) to isolate the effect of horizon on performance. The result is one of the clearest findings in the study.
What this means: The simulator’s edge is concentrated in the near-term forecast. At 1-day horizon, the generative model’s predicted return and covariance structure is accurate enough to drive a Sharpe ratio well above the equal-weight baseline. As the horizon extends, forecast accuracy degrades and the optimizer begins to react to noise rather than signal. This is an honest, expected finding — and the system surfaces it explicitly rather than hiding it.
Results — Out-of-sample stability
One good window is luck. Five consistent windows is evidence.
The critical question for any quantitative system is whether its performance is stable across different market conditions, or whether a single favorable period is doing all the work. To test this, we ran 5 overlapping rolling windows — each covering approximately 4 months of strictly post-training-cutoff data (Jan 2026 onwards), stepping forward by 2 weeks — using the optimal daily H=1 configuration.
The equal-weight and risk-parity lines track closely throughout — this is consistent with risk-parity theory, which prioritizes drawdown control over return generation. Risk parity’s value is in what doesn’t show up in a bull-run chart: its behavior during drawdowns, which our maximum drawdown metrics (consistently 9–14%) confirm is meaningfully better than an unmanaged basket.
Comparison
What “autonomous” actually requires.
The table below compares CausaQuant’s architecture against the LLM-agent-based trading systems currently being promoted in the agentic AI space. The differences are not cosmetic.
| Dimension | CausaQuant (sim + MPC) | LLM agent systems |
|---|---|---|
| Constraint satisfaction | ✓ Guaranteed — long-only, weight cap, turnover penalty enforced by the solver at every step | ✗ Approximate — LLM may violate constraints; requires validation layer |
| Human in the loop | ✓ None — after deployment. No prompt monitoring, no scoring revision, no goal redefinition | ✗ Continuous — prompts, success criteria, and scoring functions require ongoing maintenance |
| Audit trail | ✓ Full — every weight, return, turnover, and sim-to-real gap logged per rebalance period | ~ Partial — reasoning traces available but not mathematically verifiable |
| OOS validation | ✓ Rigorous — rolling windows, strict post-training-cutoff data only | ✗ Typically absent — or based on recent paper-trading with undefined methodology |
| Scalability | ✓ Parametric — add assets, adjust cadence, change horizon with no retraining | ✗ Prompt-dependent — each new asset or condition requires prompt engineering |
| Speed of decision | ✓ Milliseconds — per rebalance (solver + frozen inference) | ✗ Seconds to minutes — depending on LLM inference and tool calls |
| Regulatory explainability | ✓ Mathematical — objective function, constraint set, and optimal conditions are all written down | ✗ Black box — “the model decided” is not an acceptable compliance answer |
If you’re building with real capital, the math has to be right.
We have the simulator, the optimizer, and rolling out-of-sample evidence. If you’re a fund manager or allocator who wants to see the methodology in detail — not the marketing, the actual math — get in touch.
Request the technical brief