TradingAgents-Lightweight
A fast, cheap, momentum-first swing-trading agent. It screens for momentum stocks, scores them with a transparent rule engine, checks news and fundamentals, and returns a BUY/SELL/HOLD with an entry, an ATR stop, and a fixed 1:2 target. The LLM layer is optional — it runs free with no API keys.
English | 中文
The whole chain — data, screening, scoring, position sizing — is a few hundred lines of readable Python you can fork and adapt.
What it does
- Screens its own candidates. Auto-screens Finviz for names in a confirmed uptrend, instead of you typing tickers.
- Decides in one pass. Momentum-led technicals, a quick news long/short read, a glance at fundamentals, one call — seconds, not minutes.
- Trades with a plan. Every BUY/SELL ships with an entry, a 2×ATR stop, and a fixed 1:2 target.
- Pluggable news. SEC EDGAR 8-K filings by default (free, no key); Benzinga (paid) or any RSS feed with one flag.
- Any model, or none. OpenAI, Claude, DeepSeek, Gemini all plug in. No key? It runs the rule engine for free.
- Honest backtester. The same signal code runs live and in the backtest; equity curve and drawdown saved locally.
- Sizes and places the order. Risk-based position sizing; paper by default, Alpaca bracket orders for real.
- Bring your own alpha. The strategy is one function behind a registry —
register yours and select it with
--alpha.
Contents
- Why lightweight
- Pipeline
- How it decides
- Install
- Docker
- Auto stock picking
- Backtest
- Add your own alpha
- Place the order
- Choose your model
- Data sources
- Configuration
- Brokers
- Roadmap
- Disclaimer
Why lightweight
Full multi-agent setups (a bull researcher, a bear researcher, a risk manager, debate rounds) are great for studying how a decision gets argued out, but slow and token-heavy for actually trading. This trades the debate for speed: one pass, one decision, a couple of seconds, cheap enough to scan a whole watchlist before the open. It reads price action first, checks headlines for a long/short lean, and commits. If you want a fundamentals deep-dive, this isn't it; if you want a disciplined swing signal you can act on now, that's the idea.
Pipeline
Finviz screener ──→ momentum leaders (auto-pick) ┐
├─→ ticker(s)
or pass your own tickers ───────────────────────── ┘ │
▼
yfinance ──→ daily OHLCV
▼
momentum engine = the ALPHA ──→ score · ATR stop · 1:2 target ← swap your own
▼
news ──→ fast long/short scan (SEC · Benzinga · RSS)
▼
Finviz ──→ fundamentals: sector, P/E, market cap (context)
▼
LLM (optional) ──→ final call + one-line reason
▼
BUY / SELL / HOLD + entry · stop · target (1:2)
▼
position sizing (risk budget) ──→ broker (paper / Alpaca)
Drop the LLM and you still get a complete, deterministic decision.
How it decides
- Trade with the trend. Six checks — price vs the 20-day, 20- vs 50-day, 1- and 3-month return, MACD histogram, 20-day rate of change — each votes +1 or -1. Score ≥ +3 is a BUY, ≤ -3 is a SELL, in between is HOLD.
- Ride winners, cut losers. No bottom-fishing; an overheated RSI (> 82) blocks new longs so you're not buying the last tick of a spike.
- News is a tiebreaker, not the thesis. Headlines get a quick keyword lean; it nudges conviction, it doesn't start trades.
- Risk fixed before reward. Stop at 2×ATR, target at twice the stop distance — 1:2, always. If the math doesn't fit, it's a HOLD.
- Swing timeframe. Daily bars, holds of days to a couple of weeks, with a time stop so dead money doesn't sit.
Install
pip install tradingagents-lightweight # or: pipx install tradingagents-lightweight
pip install "tradingagents-lightweight[llm]" # with the optional LLM layer
Python 3.10+. No API key needed to start — it runs the rules for free.
talite screen # auto-pick momentum names
talite analyze NVDA --no-llm # one ticker, free
talite backtest # backtest the default basket
talite --version
Nothing to install: uvx tradingagents-lightweight screen runs it in a throwaway environment.
From source, to change the code:
git clone https://github.com/jackson-cmd/tradingagents-lightweight.git
cd tradingagents-lightweight
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e ".[llm,dev]"
Docker
docker run --rm ghcr.io/jackson-cmd/tradingagents-lightweight screen
docker run --rm -e OPENAI_API_KEY=$OPENAI_API_KEY ghcr.io/jackson-cmd/tradingagents-lightweight analyze NVDA
docker run --rm -v "$PWD/results:/app/results" ghcr.io/jackson-cmd/tradingagents-lightweight backtest
Pass keys with -e, or mount an env file with --env-file .env. To build the image yourself: docker build -t talite .
Auto stock picking
talite screen # top momentum names right now
talite screen --top 20 # cast a wider net
talite screen --backtest # screen, then backtest the picks
talite screen --top 5 --execute # screen, analyze, paper-trade the BUYs
The default filters keep it to liquid, non-penny names in a confirmed uptrend
(above the 20- and 50-day, up on the quarter), then rank by a blend of
weekly/monthly/quarterly strength. Tune the filters and weights in
talite/screener.py.
Backtest
The backtester runs the same signal code as live trading and saves an equity
curve, drawdown and trade log to results/.
talite backtest # default mega-cap basket, last month
talite backtest NVDA AMD HPE --months 3 # your basket, 3 months
talite screen --backtest # backtest the current screener picks
How it stays clean: signals on bar i only see data through bar i, positions are managed from the next bar, and a stop and target in the same bar assume the stop hit first. Sizing is fixed-fractional, so a winner is +2R and a stop-out is -1R.
A screener-picks backtest is hindsight by construction — those names are flagged because they already trended, so backtesting the month they ran flatters the result. Use it to check the loop works end to end, then paper-trade forward before trusting any number.
Add your own alpha
The strategy is one function behind a registry. Register it, select it — it trades and backtests exactly like the built-in one.
from talite.strategy import register_alpha, Signal
from talite import run_backtest
@register_alpha("gap_and_go")
def my_alpha(df, ticker="", rr=2.0, atr_mult=2.0) -> Signal:
close = float(df["Close"].iloc[-1])
# ... your edge here ...
return Signal(ticker, "BUY", close, stop=close * 0.95,
target=close * 1.10, rr=rr, momentum=0,
reasons=["gap up on volume"])
run_backtest(["NVDA", "AMD"], months=1, alpha="gap_and_go")
talite analyze NVDA --alpha gap_and_go
talite backtest NVDA AMD --alpha gap_and_go
The momentum engine ships as the default alpha ("momentum"); yours sits right
next to it — same stops, same 1:2, same backtester, same execution.
Place the order
It sizes the position so a stop-out costs exactly your risk budget (1% of equity by default), then routes it.
talite analyze HPE --execute # paper order, logged to results/
talite screen --top 5 --execute # auto-pick + paper-trade the BUYs
talite analyze HPE --execute --broker alpaca # bracket order via Alpaca (paper)
talite analyze HPE --execute --broker alpaca --live # ...for real
Paper is the default and writes to results/paper_orders.csv — nothing leaves
your machine. The Alpaca adapter places a one-shot bracket (entry + 1:2
take-profit + stop-loss) and defaults to the paper endpoint; --live is the only
thing that arms it. Sizing lives in talite/execution.py.
Choose your model
Provider-agnostic; inferred from the model name or set explicitly.
talite analyze TSLA --model gpt-4.1-mini # OpenAI
talite analyze TSLA --model claude-sonnet-4-6 # Anthropic
talite analyze TSLA --model deepseek-chat # DeepSeek
talite analyze TSLA --model gemini-2.0-flash # Gemini
gpt-4.1-mini (or -nano) is a cheap default: a short prompt in, a one-line
JSON verdict out. The LLM is optional — --no-llm runs the rules alone, free and
deterministic. The edge is in the rules; the LLM just refines the final call.
Data sources
| Layer | Source | Notes |
|---|---|---|
| Prices | yfinance | daily OHLCV, free |
| Screening + fundamentals | Finviz via finvizfinance | momentum screen, sector, P/E, market cap |
| News (free, default) | SEC EDGAR | latest 8-K filings — NEWS_SOURCE=sec |
| News (paid) | Benzinga | real-time, per-ticker — NEWS_SOURCE=benzinga |
| News (RSS) | any feed | set RSS_FEED_URL, NEWS_SOURCE=rss |
If the news layer is down, the agent falls back to price action alone and keeps going.
Configuration
Copy .env.example to .env and set only what you use:
LLM_PROVIDER=openai
LLM_MODEL=gpt-4.1-mini
OPENAI_API_KEY=sk-...
NEWS_SOURCE=sec # sec | benzinga | rss
BENZINGA_API_KEY=
SEC_USER_AGENT=tradingagents-lightweight you@example.com
RISK_REWARD=2.0
ATR_MULT=2.0
RISK_PER_TRADE=0.01
# Only if you wire live execution:
ALPACA_API_KEY=
ALPACA_SECRET_KEY=
Brokers
Paper trading works out of the box. For real fills:
- Alpaca — commission-free API with a paper
sandbox, via
alpaca-py. Wired up already (--broker alpaca). - Interactive Brokers — TWS / Client Portal API,
e.g. via
ib_async. - tastytrade — developer API and the official Python SDK.
Paper-trade first. Always.
Roadmap
- Auto stock screening (Finviz momentum)
- Pluggable alpha registry
- Pluggable news sources (SEC / Benzinga / RSS)
- Position sizing + paper / Alpaca execution
- Docker image
- More broker adapters (IBKR, tastytrade)
- Optional intraday timeframe for faster swings
- Portfolio-level sizing and exposure caps
Disclaimer
Software for research and education, not financial advice. Momentum strategies lose in chop and reversals, backtests are not the future (especially in-sample ones), and you can lose money. Trade your own account at your own risk. Paper-trade before you risk a cent.
License
MIT — see LICENSE.
Metadata
Release files for tradingagents-lightweight 0.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| tradingagents_lightweight-0.1.0.tar.gz | 52.6 kB | With provenance | Details for tradingagents_lightweight-0.1.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| tradingagents_lightweight-0.1.0-py3-none-any.whl | Python 3 | none | any | Details for tradingagents_lightweight-0.1.0-py3-none-any.whl |
Total release size: 81.9 kB
Release files / tradingagents_lightweight-0.1.0.tar.gz
| Download URL | tradingagents_lightweight-0.1.0.tar.gz |
|---|---|
| Size | 52.6 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
4f57d204446ffcd3c3dcb13e1b94ce7d5b4bf828d898b3837ac8e0920873e714
|
|
BLAKE2b-256 checksum How to use checksums |
d2063f2fb64102a6f135e90d50dd9648aeb9b8c2932ce6bfb1419ed2da933816
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 13, 2026.
Transparency logRelease files / tradingagents_lightweight-0.1.0-py3-none-any.whl
| Download URL | tradingagents_lightweight-0.1.0-py3-none-any.whl |
|---|---|
| Size | 29.3 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
389a180d817422d1cbabd7b87ca748d1d3d9bbd80f491a8d5fcbebef53142924
|
|
BLAKE2b-256 checksum How to use checksums |
44881400f2ef5867b9fa27d82c2412f50e9532c255dd9f9616ea5785cc01eeda
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 13, 2026.
Transparency logSecurity for tradingagents-lightweight 0.1.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for tradingagents_lightweight-0.1.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 13, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.