brain-topology-trader

Daily autonomous trading bot using a Neural Circuit Policy (NCP / Liquid Time-Constant) model.

Overview

Property	Value
Universe	Top ~800 US stocks by market cap
Account	Alpaca Paper Trading
Leverage	2× ($100K margin → $200K effective)
Directions	Long and Short
Model	LTC / AutoNCP (ncps PyTorch library)
Compute	Modal (Fluid Compute + A10G GPU)
State	Modal Volume trading-data at /data

Architecture

modal_app.py        ← Modal app, image, cron entrypoints
config.py           ← constants, TICKER_UNIVERSE, hyperparams
data/
  ingest.py         ← Twelve Data (8-key rotation), FRED, Finnhub
  features.py       ← TA indicators + macro + sentiment + cross-sectional rank
model/
  ncp_model.py      ← LTC + AutoNCP + stock embedding
  train.py          ← one-time 25-year supervised training
  update.py         ← daily online RL (policy gradient)
execution/
  signals.py        ← 3-day smoothing, confidence threshold, ranking
  sizing.py         ← Kelly criterion, position caps
  broker.py         ← Alpaca Paper Trading client
reward/
  compute.py        ← composite reward: direction accuracy + PnL return
utils/
  logger.py         ← structured logging
  notify.py         ← email / Telegram daily report
requirements.txt

Cron Schedule

Job	UTC	IST	ET
Inference + Execution	17:30	~23:00	13:00
EOD Weight Update	22:00	~03:30+1	17:30

Setup

1. Modal Secrets (create once)

# Alpaca Paper Trading
modal secret create alpaca-secret \
  ALPACA_API_KEY=... ALPACA_SECRET_KEY=... \
  ALPACA_BASE_URL=https://paper-api.alpaca.markets

# Twelve Data (8 keys)
modal secret create twelvedata-secret \
  TWELVE_DATA_KEY_1=... TWELVE_DATA_KEY_2=... \
  TWELVE_DATA_KEY_3=... TWELVE_DATA_KEY_4=... \
  TWELVE_DATA_KEY_5=... TWELVE_DATA_KEY_6=... \
  TWELVE_DATA_KEY_7=... TWELVE_DATA_KEY_8=...

modal secret create fred-secret FRED_API_KEY=...
modal secret create finnhub-secret FINNHUB_API_KEY=...

# Notify — pick one:
# Email:    SMTP_HOST, SMTP_PORT, SMTP_USER, SMTP_PASS, NOTIFY_EMAIL
# Telegram: TELEGRAM_BOT_TOKEN, TELEGRAM_CHAT_ID
modal secret create notify-secret NOTIFY_EMAIL=... SMTP_HOST=... SMTP_PORT=587 SMTP_USER=... SMTP_PASS=...

2. Deploy

git clone https://github.com/Rohan5commit/brain-topology-trader.git
cd brain-topology-trader
modal deploy modal_app.py

3. One-time Historical Training

modal run modal_app.py::train_historical

This fetches 25 years of OHLCV (2000–2025), trains the NCP on 3-class next-day direction labels, and saves ncp_weights_base.pt to the Modal Volume. Expect 2–6 hours on A10G.

State Files (Modal Volume /data)

File	Purpose
ncp_weights_base.pt	Base trained weights
ncp_weights_latest.pt	Daily-updated weights
signals_history.parquet	Last 3 days of raw model outputs
positions_state.parquet	Open positions + entry dates
features_cache.parquet	Latest feature snapshot

Notes

• No secrets are hardcoded anywhere — all via modal.Secret.from_name(...)
• All persistent state lives in Modal Volume, not git
• Alpaca base URL is always https://paper-api.alpaca.markets (paper trading only)