Trading Signals

NeuronLens converts hidden activations into buy/sell/hold trading signals, explains them with feature trails, and validates them with backtesting & correlations.

Outcome: move from “black-box alpha” to “auditable, regulator-ready, and monetizable trading signals.”

🔑 Core Features

Multi-asset signals: Buy / Sell / Hold across equities, FX, crypto

Backtesting environment: Historical evaluation with metrics (Sharpe, win rate, drawdown)

Feature correlations: Map signals to interpretable drivers (earnings reports, volatility, liquidity)

Explained outputs: Before/After view, linked to features like in Audit

Revenue Angle: Subscription API, licensing for funds / fintechs

🧪 Quickstart: Signal Generation + Backtest

python
from neuronlens import trading

# 1) Attach model + SAE (same as Audit)
model = trading.load_model("cxllin/Llama2-7b-Finance")
sae   = trading.load_sae(model, layers=[12,22,28])

# 2) Pick assets and timeframe
assets = ["AAPL","MSFT","BTC-USD"]
dates  = ("2024-01-01","2024-06-30")

# 3) Generate signals
signals = trading.generate_signals(
    model=model,
    sae=sae,
    assets=assets,
    start=dates[0],
    end=dates[1],
    strategy="finbert_sae_mix",   # interpretable hybrid strategy
    features_topk=5               # explain top 5 drivers per signal
)

print(signals.head())

Example output

plain text
    date        asset   signal  confidence  top_features
0  2024-01-02   AAPL    BUY     0.83        [(159,'Earnings Growth',0.41),
                                             (258,'Market Volatility',0.27)]
1  2024-01-02   MSFT    HOLD    0.55        [(345,'Performance Metrics',0.32)]
2  2024-01-02   BTC-USD SELL    0.71        [(611,'Risk Sentiment',0.38)]

🔄 Backtest in 5 lines

python
bt = trading.backtest(signals, initial_capital=100_000, metrics=["sharpe","winrate","drawdown"])
print(bt.summary())

Example output

plain text
Initial capital: $100,000
Final capital:   $127,430
Sharpe ratio:    1.62
Win rate:        57%
Max drawdown:    -8.5%

📊 Feature Correlation Explorer

python
corr = trading.feature_correlations(signals, layer=22)
print(corr[:5])

Example output

plain text
Feature 159 (Earnings Growth) → +0.43 corr with BUY signals (AAPL, MSFT)
Feature 258 (Volatility)      → +0.35 corr with SELL signals (BTC-USD)
Feature 375 (Financial Jargon)→ +0.21 corr with HOLD signals

🎨 Black-box vs Explained (Trading Edition)

Reuse the before/after API from Audit:

python
from neuronlens import audit

assets = audit.before_after_view(signals)
# shows raw vs explained signals, with feature labels

💰 Revenue Angles

Subscription API → $X/user/month for fintechs/traders

Institutional Licensing → hedge funds, prop desks, RIAs

White-label dashboards → embed in client portals

🚀 Minimal End-to-End Example

python
from neuronlens import trading

model = trading.load_model("cxllin/Llama2-7b-Finance")
sae   = trading.load_sae(model, layers=[22])

signals = trading.generate_signals(model, sae, ["AAPL"], "2024-01-01","2024-06-30")
bt = trading.backtest(signals)

print(signals.head(2))
print(bt.summary())

Output (truncated)

plain text
    date      asset signal conf
0  2024-01-02 AAPL  BUY    0.81
1  2024-01-03 AAPL  SELL   0.66

Sharpe ratio: 1.54
Final capital: $112,340