The official implementation release for our 2026 paper.
Paper | HTML | PDF | DOI | Code
Overview · Framework · Results · Agent Code · Quick Start · Reproducibility
If you use this implementation in your research, please cite TrustTrade and the upstream TradingAgents framework on which it is built:
@misc{li2026trusttrade,
title = {TrustTrade: Human-Inspired Selective Consensus Reduces Decision Uncertainty in LLM Trading Agents},
author = {Minghan Li and Rachel Gonsalves and Weiyue Li and Sunghoon Yoon and Mengyu Wang},
year = {2026},
eprint = {2603.22567},
archivePrefix = {arXiv},
primaryClass = {cs.CE},
doi = {10.48550/arXiv.2603.22567}
}Machine-readable metadata are available in CITATION.cff.
See NOTICE and docs/UPSTREAM.md for upstream
attribution.
LLM trading agents often exhibit uniform trust: retrieved facts and heterogeneous information sources are treated as equally reliable. In noisy financial environments, this can amplify hallucinations and produce unstable portfolio decisions.
TrustTrade replaces uniform trust with three human-inspired mechanisms:
- Selective consensus: multiple independent agents analyze each information domain; a credibility scorer retains corroborated semantic and numerical claims and discounts conflicting evidence.
- Deterministic temporal grounding: reproducible price-derived indicators anchor trend, momentum, volatility, drawdown and risk exposure.
- Reflective memory: short- and long-horizon return/risk feedback adapts information weighting and risk preferences at test time, without training.
The paper evaluates AAPL, GOOG and NVDA, across 2024 Q1 historical backtesting and a 2026 Q1 forward-time study, with 19 human annotators as a behavioral reference.
Each decision moves through a controlled evidence pipeline:
Ticker + date + portfolio
│
▼
Deterministic temporal signals
│
▼
Parallel domain analysts ──► credibility scoring ──► high-consensus evidence
│
▼
bull/bear researchers
│
▼
BUY / HOLD / SELL + %
│
▼
short/long reflective memory
- Naive multi-source aggregation is not additive. In the 2024-Q1 study, market-only signals at the Analyst stage reach 33.1% cumulative return at 9.1% maximum drawdown, while indiscriminately adding sources can inject noise rather than improve decisions.
- Large LLM traders occupy an aggressive regime. Full-allocation agents achieve roughly 30% average cumulative return, but incur approximately 12% maximum drawdown.
- Selective consensus stabilizes intermediate decisions. The high-consensus configuration raises average cumulative return from roughly 10% to 26%, while moving maximum drawdown from about 3% to 8%.
- Temporal signals add reproducible grounding. They improve average return by about 1 percentage point while slightly reducing drawdown.
- Memory and reflection regularize risk. They trade a small amount of return for lower drawdown, moving the operating point toward the human-aligned mid-risk/mid-return region.
| Evaluation | Universe | Window | Purpose |
|---|---|---|---|
| Historical controlled study | AAPL, GOOG, NVDA | 2024 Q1 | diagnose source, reasoning-depth and allocation effects |
| Human comparison | 19 annotators | 2024 Q1 scenarios | characterize selective information use and risk preference |
| Forward-time backtest | AAPL, GOOG, NVDA | 2026-01-01 to 2026-02-18 | reduce future-information leakage |
TrustTrade shifts partial-allocation base models toward a stronger return–risk frontier. Adding memory and reflection produces the more conservative variant. The shaded region is the human-aligned reference defined in the paper.
Note
These values summarize the paper experiments; they are not a promise of live trading performance. Backtested performance does not guarantee future results.
The complete lmh research-branch agent package is present: 44 of 44 Python
source files under tradingagents/ are included. Public-release changes are
limited to credential safety, portable configuration, dependency-compatible
imports and defensive bug fixes.
| Stage | Core implementation | Responsibility |
|---|---|---|
| Temporal analyst | price_analyst.py |
deterministic price history, trends and forecasts |
| Domain analysts | agents/analysts/ |
fundamentals, market, news and social reports |
| Credibility layer | agents/credibility_scorer/ |
cross-agent consistency, confidence filtering and leakage audit |
| Research debate | agents/researchers/ |
bull/bear evidence synthesis |
| Trader | trader.py |
final action and position sizing |
| Reflective memory | memory_bank.py and reflection.py |
short/long-horizon outcome feedback |
| Orchestration | trading_graph.py and setup_multiagent.py |
graph construction and multi-agent execution |
| Optional risk debate | agents/risk_mgmt/ |
upstream-compatible aggressive/neutral/conservative debate |
The paper's main experiments omit the final Risk Manager stage after finding no systematic improvement, but its implementation remains available for ablations and upstream compatibility.
tradingagents/
agents/ complete analyst, scorer, researcher and trader stack
graph/ orchestration, consensus, memory and reflection
dataflows/ online/offline data interfaces
baseline_analysis/ rule-based and LLM backtesting code
configs/
models.template.json provider/model registry without credentials
experiments/ safe example experiment manifests
tools/ plotting code and consensus diagrams
assets/ README framework and result figures
docs/
DATA.md data, privacy and redistribution policy
RELEASE_CHECKLIST.md security-to-publication author checklist
REPRODUCIBILITY.md reproduction workflow and known boundaries
UPSTREAM.md TradingAgents provenance and modifications
scripts/
check_release.py offline secret/privacy/structure checks
run_experiment.py manifest-driven experiment launcher
validate_release.sh one-command validation; no API calls
tests/ offline release and configuration tests
This public tree intentionally excludes credentials, notebook state, person-level human-study records, generated run directories and provider-restricted data caches.
git clone https://github.com/Harvard-AI-and-Robotics-Lab/TradingAgent.git
cd TradingAgent
python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
pip install -e .
bash scripts/validate_release.shFor rule-based baselines and plotting utilities, install the analysis extras:
pip install -e ".[analysis]"The validation command checks Python syntax, required release files, JSON manifests, accidental credentials, private machine paths and person-level data fields. It makes no network requests or model API calls.
Copy .env.example only as a reference and export credentials through your
shell or secret manager. Never commit .env files or notebook credentials.
export OPENAI_API_KEY=...
export FINNHUB_API_KEY=...
# Optional providers
export ANTHROPIC_API_KEY=...
export GOOGLE_API_KEY=...
export XAI_API_KEY=...Provider availability and model identifiers change over time. Review
configs/models.template.json before running an
experiment.
The default trusttrade mode uses two independent analyst calls and a
credibility scorer. This is a small functional example, not the full paper
sweep.
python main.py \
--ticker NVDA \
--trade-date 2024-05-10 \
--mode trusttrade \
--provider openai \
--model gpt-4o-miniUse --mode base for the upstream single-analyst path. Run
python main.py --help for multi-provider analyst and scorer options.
Preview an experiment command without making API calls:
python scripts/run_experiment.py \
--config configs/experiments/2024_q1.example.json \
--symbol AAPL \
--dry-runRemove --dry-run to execute.
Important
TrustTrade can make multiple paid model and data-provider calls per trading decision. Review provider pricing and rate limits before running a sweep.
| Level | Status | What it means |
|---|---|---|
| Offline software validation | ✅ | syntax, manifests, release hygiene and tests pass without APIs |
| Clean-environment installation | ✅ | package, core imports and both CLI entry points are verified |
| Functional provider run | 🔑 | requires valid model/data credentials and incurs external cost |
| Exact paper-result reproduction | 🚧 | requires author-approved frozen manifests, caches and model snapshots |
The supplied *.example.json manifests are safe launch templates; they are not
represented as the exact paper manifests. See
docs/REPRODUCIBILITY.md for the scientific
reproduction contract and docs/RELEASE_CHECKLIST.md
before making the repository public.
This repository does not distribute identifiable participant records. Human
study artifacts should be released only after a separate consent, ethics and
de-identification review. Market, news, social-media and model-generated data
may carry provider-specific redistribution restrictions. See
docs/DATA.md.
TrustTrade is derived from
TauricResearch/TradingAgents,
released under Apache-2.0. TrustTrade adds multi-agent credibility scoring,
selective-consensus routing, deterministic temporal signals, portfolio-aware
backtesting and reflective memory. Please retain upstream attribution when
redistributing this work.
This software is for research and educational purposes only. It is not financial, investment, legal or trading advice. Users are responsible for provider costs, data rights, model outputs and any decisions made with this software.
The code is released under the Apache License 2.0. Third-party data, model outputs and service APIs are governed by their respective terms.