01 ROLE

AI STRATEGY RESEARCH LAB

Use GPT to discover and improve the method. Let deterministic code execute it.

AlgoTrading Lab separates research intelligence from runtime execution. The LLM can inspect charts, formulate a trading hypothesis, generate strategy.py, analyze backtest failures and optimize the next version. Once approved, the strategy is judged and executed by the deterministic engine rather than by an LLM decision on each candle.

EXAMPLE RESEARCH CASE Find one common setup across several instruments and reduce drawdown without killing trade activity.

The model can compare successful and failing trades, challenge the original hypothesis and write the next deterministic version. The backtest result becomes evidence for the next research step.

02 PIPELINE

HOW THE LAB WORKS

Data → hypothesis → code → deterministic result → hypothesis correction.

MARKET EVIDENCE Charts, instruments and the research question

Start with the market behavior you want to investigate. The goal can be discovery, debugging, lower drawdown, smoother returns or understanding why a strategy fails on specific instruments.

RESEARCH GPT studies patterns and failure examples

Builder Research can compare chart structures, regimes and previous backtest evidence before committing to deterministic logic.

HYPOTHESIS → CODE Turn the method into strategy.py

Once the hypothesis is explicit, GPT generates or repairs deterministic strategy code inside the project contract. The strategy artifact is inspectable and versioned.

BACKTEST The deterministic engine becomes the judge

Historical candles are replayed through the frozen strategy code. Trade count, PnL, drawdown, instrument-level results and the equity curve show what the hypothesis actually produced.

ERROR ANALYSIS Feed the full result back into research

The model analyzes where the hypothesis failed, whether the error is structural or parameter-like, and whether the next step should refine the same method or return to discovery.

OPTIMIZATION Improve within the strategy thesis

Candidate versions are generated and judged by full deterministic backtests. The result is fed back into the next optimization step rather than being optimized from an isolated metric.

RUNTIME Approved code runs deterministically

Paper and live execution use the approved strategy.py without an LLM call on every candle. GPT returns when you want to research, repair, explain or optimize the strategy again.

03 USE CASES

EXAMPLE RESEARCH WORKFLOWS

Use the LLM as a researcher and debugger around deterministic evidence.

01

DISCOVERY

Find one common setup across instruments

Ask Builder Research to compare multiple chart series and identify a structure that can be expressed as one universal deterministic method.

OUTPUT: hypothesis → strategy.py → instrument-level backtest.

02

CURVE QUALITY

Flatten the equity curve

Give the model the full backtest and ask which parts of the hypothesis create clustered losses, unstable returns or excessive drawdown.

OUTPUT: revised logic judged against the complete previous result.

03

FAILURE ANALYSIS

Understand why specific pairs fail

Compare profitable and defective instruments, inspect their market structure and determine whether the issue is a missing regime filter or a bad universal assumption.

OUTPUT: targeted hypothesis correction rather than blind parameter search.

04

ACTIVITY VS RISK

Reduce drawdown without killing activity

Optimize around the existing strategy thesis while preserving trade retention and checking whether improvements generalize across instruments.

OUTPUT: a new deterministic version with measurable trade-offs.

04 STRUCTURE

LAB STRUCTURE

GPT does the research work. The deterministic engine validates and executes the result.

01 Chat + Builder Research Direct the research, inspect market evidence, ask for alternative hypotheses and analyze backtest failures.

02 strategy.py The deterministic strategy artifact generated from the current hypothesis and kept inside the project version history.

03 Deterministic Backtest Historical replay executes the frozen code and produces the evidence used to judge the strategy thesis.

04 Optimization + Runtime GPT improves candidate logic; accepted strategy code then runs deterministically in paper or live execution.

RESEARCH LAYER Charts · hypotheses · backtest evidence · Chat

GPT BUILDER Research · strategy generation · repair · optimization

DETERMINISTIC ENGINE strategy.py · backtest · paper · live

AVAILABLE GPT TIERS Use stronger models where the research problem needs them.

STANDARD

GPT-5.4 mini

Fast, lower-cost model for routine Chat, code changes and economical research cycles.

Input: $0.98 / 1M
Cached: $0.10 / 1M
Output: $5.85 / 1M

RECOMMENDED

GPT-5.6 Terra

Stronger reasoning with a practical intelligence-to-cost balance for research and strategy development.

Input: $2.60 / 1M
Cached: $0.26 / 1M
Output: $15.60 / 1M

FRONTIER

GPT-5.6 Sol

Highest reasoning tier for difficult strategy discovery, debugging and multi-step optimization work.

Input: $6.50 / 1M
Cached: $0.65 / 1M
Output: $39.00 / 1M

ILLUSTRATIVE COST AI cost stays in research and code work — deterministic replay and runtime do not call GPT per candle.

A representative workflow with 10 Chat turns, one strategy compile and five LLM optimization stages is approximately $0.37 on GPT-5.4 mini, $0.99 on GPT-5.6 Terra or $2.47 on GPT-5.6 Sol. Deterministic backtests and approved paper/live runtime add $0 LLM usage. This is an illustrative workflow, not a fixed quote.

01Top up Art Argentum balance

02Use GPT for research / generation

03Run deterministic backtests at $0 LLM

04Approved runtime stays deterministic

05 START

OPEN ALGOTRADING LAB

Start with market evidence. End with testable deterministic code.

Use GPT as researcher, strategist and debugger, while the deterministic engine remains the judge of every strategy version and the executor of the approved method.

Open AlgoTrading Lab