Can Agentic Trading Systems Pay for Their Own Intelligence?

Qiqi Duan^1,*, Changlun Li^1,2,*, Chen Wang^1,*, Fan Zhang^2,4, Mengxiang Wang¹, Dayi Miao¹, Peixian Ma², Jiangpeng Yan³, Liyuan Chen³, Shuoling Liu³, Preslav Nakov⁴, Yuyu Luo^1,2, Nan Tang^1,2

¹HKUST(GZ) ²Paradoox AI Research ³E Fund Management Co., Ltd ⁴MBZUAI

^*Equal contribution.    Contact: nantang@hkust-gz.edu.cn

🔥 Our work is accepted by EMNLP 2026 Findings. Welcome to star and cite our work! ✨

---

📖 Overview

TradeLens is a trace-grounded toolkit for measuring and diagnosing the profit vs. cost of agentic trading systems.

Large language model (LLM) agents are increasingly used in trading systems, where reasoning, tool use, and repeated portfolio decisions introduce non-negligible deployment costs. A system can look profitable on returns yet still fail to create useful value once costs are accounted for, and diagnosing the root cause is hard because profit and cost arise from intertwined mechanisms.

We argue that agentic viability should be a central evaluation criterion: dynamic LLM-mediated decisions should generate enough incremental profit to justify the costs they induce. TradeLens reconstructs trading trajectories from records and traces, attributes profit and cost to interpretable evidence, and diagnoses whether, where, and why an agent succeeds or fails to pay for its own intelligence.

Overall, TradeLens supports:

1. Measurement — profit and cost attribution for multi-factor agentic trading deployments.
2. Diagnosis — trading performance analysis and system-level diagnosis with evidence-grounded revision suggestions.

📚 Citations

@inproceedings{duan2026can,
  title={Can Agentic Trading Systems Pay for Their Own Intelligence?},
  author={Duan, Qiqi and Li, Changlun and Wang, Chen and Zhang, Fan and Wang, Mengxiang and Miao, Dayi and Ma, Peixian and Yan, Jiangpeng and Chen, Liyuan and Liu, Shuoling and Nakov, Preslav and Luo, Yuyu and Tang, Nan},
  booktitle={The 2026 Conference on Empirical Methods in Natural Language Processing},
  year={2026},
  url={https://openreview.net/forum?id=4AWZckdl2J}
}

📑 Introduction

Repository Structure

The toolkit loads experiment records and static config from data/, computes trading costs (commission, token usage, infra, monthly data subscription, and a per-day uncertain add-on), reconstructs portfolio value, and decomposes profit into market / selection / timing effects (full-cash and invested-only benchmarks).

It writes reports under result/ and, when configured, runs LLM diagnosis on top of the financial report.

TradeLens/
├── 📁 data/              # Input: records, static configs, price JSONL
├── 📁 toolkit/           # Toolkit: measurement + diagnosis
│   ├── 📄 main.py        # CLI and end-to-end pipeline
│   ├── 📁 cost/          # Commission, token/infra/monthly/uncertain costs
│   ├── 📁 profit/        # Records, portfolio valuation, profit breakdown
│   ├── 📁 report/        # Financial report, plots, LLM agent, HTML render
│   └── 📄 config.py      # Load config.json, static config, LLM pricing
├── 📁 result/            # Output: one folder per run (gitignored)
├── 📄 main.py            # Entry: python main.py → toolkit.main
├── 📄 config_llm.json    # Per-model token pricing
├── 📄 config.json        # Paths into data diagnosis settings
└── 📄 env.example        # API keys for LLM diagnosis

Documentation Navigation

• 📖 Overview
• 📚 Citations
• 📑 Introduction
  • Repository Structure
  • Documentation Navigation
• 📝 Usage
  • Quick Start
  • CLI
  • Inputs (data/ + configuration)
  • Outputs (result/)
• 🔌 Integrate with your agentic trading system
  • Step 1. Export experiment_records.jsonl
  • Step 2. Record your setting to static.json
  • Step 3. Provide prices data (price-data.jsonl)
  • Step 4. Run TradeLens on your exported artifacts
• 🔧 Troubleshooting

📝 Usage

Quick Start

1. python -m venv .venv && source .venv/bin/activate
2. pip install -r requirements.txt
3. Edit config.json, and point records_path / static_path to your files under data/
4. Edit config_llm.json so each model in your records has a pricing row (otherwise token cost will be 0)
5. Copy .env.example' to .env' and set your URL and API key
6. Run:
   • python main.py
   • (optional overrides) python main.py --records data/...jsonl --static data/...jsonl --out result/my-run

CLI

python main.py --help
python main.py [--config CONFIG] [--records RECORDS] [--static STATIC] [--out OUT] [--benchmark SYMBOL]

Inputs (data/ + configuration)

configuration

Field	Description
records_path	Experiment records JSONL, e.g. data/experiment_records_*.jsonl
static_path	Static config JSON/JSONL, e.g. data/static-*.jsonl
prices_path	Daily prices JSONL, e.g. data/price-data.jsonl
market_base_path	Optional benchmark price JSONL, e.g. data/market-base.jsonl
agent_model	LLM model for diagnosis (optional)
benchmark_symbol	Benchmark symbol (default SPY)
llm_call_success_rate	Scales token costs (default 1.0)

Record / static / price conventions (high level)

• Records (JSONL): each line is a trading-day record; key fields include date, llm_usage, trades (with decision_type, ticker, quantity, and prices).
• Static (JSON or JSONL): deployment metadata such as initial_cash, decision_frequency, data_subscription_monthly, start_time, end_time, and llm_model.
• Prices (JSONL): Alpha Vantage–style daily buy prices per symbol (see data/price-data*.jsonl).

Outputs (result/)

Default output directory:

• result/<llm_model>-<initial_cash>-<frequency>/ (or pass --out)

Common files:

• *-action.txt: daily actions and cost ledger
• *-financial-report.md and *-financial-report.html: financial report + embedded charts
• *-llm-analysis.md and *-llm-analysis.html: LLM diagnosis (when enabled)
• charts/*.png: charts referenced by the markdown reports

🔌 Integrate with your agentic trading system

TradeLens is an after-the-fact analysis toolkit: you run your own agentic trading system, export its records/traces, then run TradeLens to attribute profit & cost and generate reports.

Step 1. Export experiment_records.jsonl

Write one JSON object per line (JSONL). Each line corresponds to one decision day (recommended date format: YYYY-MM-DD).

Minimal example:

{"date":"2026-01-02","model":"gpt-5.2","llm_usage":{"model":"gpt-5.2","input_tokens":1200,"output_tokens":450,"cached_tokens":0,"latency_ms":820},"trades":[{"decision_type":"BUY","ticker":"AAPL","quantity":10,"analysis_price":188.12,"execution_price":188.40,"timestamp":{"analysis_time":"2026-01-02T14:30:00Z","decision_time":"2026-01-02T14:30:01Z"}}]}

Notes:

• llm_usage.model must match a key in config_llm.json (otherwise token cost cannot be priced).
• If you provide both analysis_price and execution_price, TradeLens can estimate slippage/opportunity cost.
• If you provide timestamp.analysis_time and timestamp.decision_time, TradeLens can compute analysis→decision latency.

Step 2. Record your setting to static.json

TradeLens expects a JSON file with a required structure field. Fields in structure are merged into the top-level config.

Minimal example:

{
  "structure": {
    "llm_model": "gpt-5.2",
    "initial_cash": 100000,
    "decision_frequency": "daily",
    "data_subscription_monthly": 100.0,
    "start_time": "2026-01-01",
    "end_time": "2026-01-30"
  }
}

Step 3. Provide prices data (price-data.jsonl)

Provide trading asset prices for your predefined time window from the same data source as your system.

{"Meta Data":{"2. Symbol":"AAPL"},"Time Series (Daily)":{"2026-01-02":{"1. buy price":"188.12"},"2026-01-03":{"1. buy price":"190.05"}}}

Step 4. Run TradeLens on your exported artifacts

python main.py --records /path/to/experiment_records.jsonl --static /path/to/static.json --out result/my-run

If your system makes multiple intraday decisions, pre-aggregate records to hourly entries.

🔧 Troubleshooting

• records_path not found / static_path not found: ensure config.json points to existing files under data/, or pass --records/--static explicitly.
• Profit breakdown warning about BUY/SELL: if your records contain only HOLD days (no executed trades), profit breakdown charts/attribution may be skipped.
• Plotting warnings (matplotlib / numpy): reinstall dependencies in a fresh venv (e.g. recreate .venv, then pip install -r requirements.txt).
• Uncertain cost changes between runs: the per-day uncertain add-on in toolkit/main.py has no default seed. Call random.seed(<int>) at process start (or immediately before that line) if you want comparable reports across runs.