project / June 3, 2026 / 7 min read
Systematic Trading Research
A local-first strategy research framework for turning human hypotheses into reproducible MT5 experiments.
Problem
Creating systematic trading portfolios as a solo retail trader is mostly a research-throughput problem. The hard parts are not only strategy ideas, but data processing, repeatable experimentation, testing, implementation quality, and knowing which results deserve attention.
Small professional shops still have advantages in research workflow, data quality, portfolio construction, risk management, execution modeling, software engineering discipline, expensive datasets, and capital size. The gap is not closed, but parts of it are narrowing. Agentic AI raises research throughput, lowers the software engineering barrier, and makes it easier to learn from public technical writing without turning every idea into a slow manual build.
At the same time, a small retail trader has structural advantages: lower costs, less capital to deploy, lower regulatory overhead, and the ability to pursue small uncorrelated edges that are too capacity-constrained or inefficient for larger firms. The project is about building the operating system around that reality.
Approach
The workflow keeps the creative and judgment-heavy work with the human. I define the hypothesis, trading logic, target behavior, and interpretation criteria. Agents translate accepted ideas into MQL5, run compile checks, execute MT5 backtests, summarize evidence, and present the runs that are most aligned with the target.
The current emphasis is on stacking small edges, especially on higher timeframes, and using automation to test variations quickly without losing the reason each test was run. For example, when a strategy needs consolidation filters, the target may be lower frequency, lower drawdown, and a better Calmar ratio rather than simply a higher net profit.
Architecture
- Structured hypothesis intake captures assumptions, target market conditions, invalidation criteria, and evidence requirements.
- Implementation requests turn approved hypotheses into MQL5 tasks, test plans, and reproducible experiment definitions.
- Isolated MT5 workers run compile checks, one-off backtests, sweeps, and validation gates.
- Evidence reports preserve run summaries, trade data, equity curves, logs, and result dashboards.
- A review layer tracks experiment lineage, pass/fail gates, robustness notes, and promotion decisions.
The private workflow can currently run two MT5 workers in parallel. A single backtest can complete in seconds, and larger sweeps across assets, filters, and parameter variations can produce hundreds of runs in minutes. The useful part is not raw speed by itself; it is that each result can be tied back to a hypothesis and compared against a stable baseline.
Lessons Learned
The fully agentic pipeline works, but it is not yet better than human research. Autonomous strategy discovery still tends toward weak assumptions, overfit ideas, or tests that look plausible without being strategically meaningful.
Used as an exoskeleton, though, the system is already a speed multiplier. Agents accelerate coding, testing, summarization, and presentation. The human still owns the creative direction, skepticism, and interpretation.
The long-term value is cumulative. Each experiment adds to a memory of what approaches worked, where they failed, and which filters or design choices solved specific problems. That knowledge can then be reused in future strategies instead of disappearing into isolated backtest folders.
Future Work
The next step is making the evidence gates stricter and the agent workflow more selective. The system should become better at detecting contaminated tests, identifying robustness problems, comparing runs against the right baseline, and explaining why a result does or does not support the original hypothesis.