Alternative Data and NLP in Quant Strategies › Key Parameters

import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split

# Simulated financial news dataset with labeled sentiment
# In practice, this would come from labeled datasets or proxy labels
financial_headlines = [
    # Positive headlines
    ("Apple reports record quarterly revenue, beats analyst expectations", 1),
    ("Strong job growth signals economic expansion", 1),
    ("Company announces dividend increase and share buyback program", 1),
    ("Merger creates synergies expected to boost earnings", 1),
    ("New product launch exceeds sales targets", 1),
    ("Upgraded to buy rating on improved outlook", 1),
    ("Revenue growth accelerates amid strong demand", 1),
    ("Profit margins expand on cost efficiencies", 1),
    ("Company wins major contract worth billions", 1),
    ("Earnings beat estimates for fifth consecutive quarter", 1),
    # Negative headlines
    ("Company misses earnings estimates, shares plunge", -1),
    ("Regulators launch investigation into accounting practices", -1),
    ("CEO resigns amid declining performance", -1),
    ("Quarterly loss widens on falling demand", -1),
    ("Company announces layoffs as sales decline", -1),
    ("Downgraded to sell on deteriorating fundamentals", -1),
    ("Supply chain disruptions hurt profit margins", -1),
    ("Company faces lawsuit over product defects", -1),
    ("Revenue forecast cut due to weak demand", -1),
    ("Credit rating downgraded on rising debt levels", -1),
    # Neutral headlines
    ("Company reports quarterly results in line with expectations", 0),
    ("Management provides guidance for next fiscal year", 0),
    ("Annual shareholder meeting scheduled for May", 0),
    ("Company announces leadership transition", 0),
    ("Quarterly revenue unchanged from prior year", 0),
    ("Analyst maintains hold rating on stock", 0),
]

# Expand dataset by creating variations
np.random.seed(42)
texts = [h[0] for h in financial_headlines]
labels = [h[1] for h in financial_headlines]

# Convert to TF-IDF features
tfidf = TfidfVectorizer(max_features=100, ngram_range=(1, 2))
X = tfidf.fit_transform(texts)
y = np.array(labels)

# Split data
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.3, random_state=42, stratify=y
)

# Train logistic regression classifier
model = LogisticRegression(max_iter=1000, random_state=42)
model.fit(X_train, y_train)

# Evaluate
y_pred = model.predict(X_test)
y_pred_proba = model.predict_proba(X_test)

import numpy as np

# Get feature importances from logistic regression coefficients
feature_names = tfidf.get_feature_names_out()
coefficients = model.coef_

# For multiclass, we have coefficients for each class
# Class order: -1 (negative), 0 (neutral), 1 (positive)

# Extract top features for each class
top_features_by_class = {}
for class_idx, class_name in enumerate(["Negative", "Neutral", "Positive"]):
    coef = coefficients[class_idx]
    top_indices = np.argsort(coef)[-5:][::-1]
    top_features_by_class[class_name] = [
        (feature_names[idx], coef[idx]) for idx in top_indices if coef[idx] > 0
    ]

# Note: In production, you would use actual FinBERT or similar models
# This demonstrates the interface pattern

# Example using transformers library (conceptual - requires model download)
# from transformers import AutoTokenizer, AutoModelForSequenceClassification
# import torch

# For demonstration, we'll create a mock FinBERT-like interface
class MockFinBERT:
    """Simulates FinBERT sentiment analysis for demonstration."""

    def __init__(self):
        # Keywords that influence sentiment
        self.positive_signals = [
            "beat",
            "exceed",
            "strong",
            "growth",
            "profit",
            "upgrade",
            "outperform",
            "record",
            "surge",
        ]
        self.negative_signals = [
            "miss",
            "decline",
            "weak",
            "loss",
            "downgrade",
            "disappoint",
            "fall",
            "cut",
            "concern",
            "risk",
        ]
        self.negation_words = [
            "not",
            "no",
            "never",
            "neither",
            "without",
            "lack",
        ]

    def analyze(self, text):
        """Return sentiment probabilities for text."""
        words = text.lower().split()

        # Check for negation
        has_negation = any(w in self.negation_words for w in words)

        pos_count = sum(
            1 for w in words if any(p in w for p in self.positive_signals)
        )
        neg_count = sum(
            1 for w in words if any(n in w for n in self.negative_signals)
        )

        # Simulate probability distribution
        if pos_count > neg_count:
            base_probs = [0.1, 0.2, 0.7]
        elif neg_count > pos_count:
            base_probs = [0.7, 0.2, 0.1]
        else:
            base_probs = [0.25, 0.5, 0.25]

        # Negation flips positive and negative
        if has_negation:
            base_probs = [base_probs[2], base_probs[1], base_probs[0]]

        return {
            "negative": base_probs[0],
            "neutral": base_probs[1],
            "positive": base_probs[2],
        }


finbert = MockFinBERT()

# Test cases that highlight deep learning advantages
test_cases = [
    "Earnings exceeded expectations significantly",
    "Earnings did not meet expectations",  # Negation handling
    "The company reported a surprising lack of growth",  # Complex negation
    "Despite challenges, revenue remained strong",  # Contrast handling
]