Inflation Sentiment & Rates Expectations NLP Pipeline

A quantitative macro research project developed for the Citadel Central Bank Challenge, where our team finished 2nd place. The project combines natural language processing, macroeconomic analysis, and financial market interpretation to generate a forward-looking inflation and rates sentiment signal from news data.

The core objective was to support a monetary policy recommendation by extracting market-relevant sentiment from financial news and central bank commentary, then aggregating it into interpretable daily macro signals.

---

🏆 Competition Result

Field	Detail
Competition	Citadel Central Bank Challenge
Placement	2nd Place
Prize	Narrowly missed the £15,000 team prize

Our team's final recommendation was to hold interest rates, while highlighting that future hikes may be necessary if inflation expectations show signs of de-anchoring.

---

Motivation

Traditional macroeconomic indicators (CPI, wage growth, services inflation, central bank communications) are released with delays and often require interpretation. Financial news provides a high-frequency source of information about:

• Inflation expectations
• Interest rate expectations
• Monetary policy sentiment
• Market reaction to macroeconomic releases
• Forward-looking commentary from analysts and policymakers

The goal of this project was to create a systematic way of extracting that information from text and converting it into a usable macro signal.

---

Key Features

• Automated financial news collection
• Trusted-domain filtering for source reliability
• Sentence-level macro relevance filtering
• Forward-looking keyword detection
• Transformer-based sentiment classification
• Directional inflation sentiment scoring
• Entropy-based confidence weighting
• Article-level and daily aggregation
• Rolling averages and smoothing
• Data visualisation for policy presentation
• Interpretation of NLP outputs in a macroeconomic framework

---

Methodology

1. Data Collection

The pipeline collects financial news articles from selected trusted sources using news APIs. Sources are restricted to high-quality macroeconomic, financial, and institutional domains to reduce noise and improve signal reliability.

Source categories include: central banks, financial newspapers, market data providers, and reputable economic journalism outlets.

---

2. Article Filtering

Raw articles are filtered using macro-relevant keywords such as:

inflation · CPI · core inflation · interest rates · rate hikes · rate cuts · monetary policy · Bank of England · Federal Reserve · ECB · services inflation · wage growth · sticky inflation · disinflation

---

3. Sentence Extraction

Rather than classifying entire articles, the pipeline extracts individual sentences most likely to contain forward-looking macro information. This improves signal quality by isolating directionally relevant text from historical context or political commentary.

Forward-looking terms prioritised:

expected · forecast · likely · projected · outlook · anticipate · may rise · could fall · risk of · pricing in

---

4. Sentiment Classification

Each selected sentence is passed through a transformer-based classifier predicting whether inflation or rate expectations are likely to move Up, Neutral, or Down.

The core directional score is:

score = P(up) - P(down)

This creates a continuous sentiment score rather than relying on hard class labels.

---

5. Confidence Weighting

Entropy-based confidence weighting is applied so that uncertain predictions contribute less to the final signal.

Entropy:

H = -Σ p_i · log(p_i)

Normalised entropy:

H_norm = H / log(3)

Sentence confidence weight:

w = |P(up) - P(down)| × (1 - H_norm)

This rewards predictions that are both directionally strong and low-entropy.

---

6. Article-Level Aggregation

Each article receives a weighted sentiment score from its extracted sentences:

article_score = Σ(w_i × s_i) / Σ(w_i)

---

7. Daily Aggregation

Article scores are aggregated by date to produce a daily macro sentiment time series using simple averages, confidence-weighted averages, rolling averages, or exponential smoothing.

Signal interpretation:

Signal	Interpretation
Positive	Upward pressure on inflation / rates expectations
Near-zero	Neutral or mixed expectations
Negative	Downward pressure on inflation / rates expectations

---

Project Structure

.
├── data/
│   ├── raw/
│   │   └── raw_articles.csv
│   ├── processed/
│   │   └── daily_news_sentiment_results.csv
│   └── external/
│       └── macro_event_dates.csv
│
├── notebooks/
│   ├── 01_data_collection.ipynb
│   ├── 02_sentence_filtering.ipynb
│   ├── 03_sentiment_classification.ipynb
│   ├── 04_signal_aggregation.ipynb
│   └── 05_visualisation_and_analysis.ipynb
│
├── src/
│   ├── data_collection.py
│   ├── article_filtering.py
│   ├── sentence_extraction.py
│   ├── sentiment_model.py
│   ├── aggregation.py
│   ├── entropy.py
│   ├── plotting.py
│   └── utils.py
│
├── outputs/
│   ├── figures/
│   └── results/
│
├── requirements.txt
├── README.md
└── .gitignore

---

Core Components

Module	Responsibilities
data_collection.py	Query construction, API requests, date-based collection, source filtering
article_filtering.py	Removing irrelevant articles, trusted domain prioritisation, topic filtering
sentence_extraction.py	Sentence tokenisation, forward-looking detection, relevance scoring
sentiment_model.py	Transformer / SetFit inference, class probabilities, directional scores
entropy.py	Prediction entropy, normalised entropy, confidence weighting
aggregation.py	Sentence → article → daily signal aggregation, rolling averages
plotting.py	Daily sentiment charts, rolling averages, CPI / BoE event overlays

---

Signal Formulae

Sentence score:

s_i = P_i(up) - P_i(down)

Sentence confidence weight:

w_i = |P_i(up) - P_i(down)| × (1 - H_i / log(3))

Article sentiment score:

S_article = Σ(w_i · s_i) / Σ(w_i)

Daily sentiment score:

S_day = mean(S_article)

Confidence-weighted daily score:

S_day_weighted = Σ(W_j · S_j) / Σ(W_j)

Where W_j is the confidence weight for article j.

---

Visual Outputs

• Daily inflation sentiment score
• Confidence-weighted daily sentiment score
• Rolling sentiment averages
• Article count by day
• Inter-article disagreement / entropy
• Sentiment overlaid with CPI release dates
• Sentiment overlaid with Bank of England meeting dates

---

Installation

# Clone the repository
git clone https://github.com/your-username/inflation-sentiment-nlp.git
cd inflation-sentiment-nlp

# Create and activate a virtual environment
python -m venv venv
source venv/bin/activate        # macOS/Linux
venv\Scripts\activate           # Windows

# Install dependencies
pip install -r requirements.txt

---

Example Usage

Collect articles:

from src.data_collection import build_daily_news_results

results = build_daily_news_results(
    api_key="YOUR_API_KEY",
    query=query,
    start_date="2026-03-01",
    end_date="2026-04-01",
    domains=trusted_domains
)

Run sentiment classification:

from src.sentiment_model import rates_expectations

output = rates_expectations(
    model=model,
    texts=selected_sentences,
    labels=("down", "neutral", "up")
)

Aggregate daily scores:

from src.aggregation import aggregate_daily_sentiment

daily_results = aggregate_daily_sentiment(article_scores)

Plot the final signal:

from src.plotting import plot_daily_sentiment

plot_daily_sentiment(daily_results)

---

Technologies Used

pandas · NumPy · matplotlib · scikit-learn · Hugging Face Transformers · SetFit · sentence-transformers · NewsAPI / GNews · trafilatura

---

Limitations

• News coverage may be biased toward major economies
• Global financial news may not perfectly represent UK-specific inflation expectations
• API limits can restrict article coverage
• Text classifiers can misread sarcasm, hedging, or complex policy language
• Short-term news sentiment may be noisy
• Sentiment does not directly imply causation

The pipeline should be interpreted as a supporting quantitative signal, not a standalone decision engine.

---

Future Improvements

• Add Twitter / X data for higher-frequency market sentiment
• Incorporate central bank speeches directly
• Compare signal against market-implied rates curves
• Test correlation with gilt yields and SONIA futures
• Build a UK-specific article source filter
• Fine-tune a domain-specific macro-finance language model
• Add topic modelling to separate inflation, growth, labour, and policy themes
• Build a real-time monitoring dashboard
• Backtest signal behaviour around historical CPI and BoE events

---

Key Takeaways

This project demonstrated that NLP can convert unstructured financial news into a structured macroeconomic signal. The most important design choice was carefully controlling the full pipeline:

source quality → sentence relevance → forward-looking filtering → probabilistic classification → confidence weighting → macro interpretation

The result is a robust, interpretable signal capable of supporting policy reasoning.

---

Authors

Developed as part of a team submission for the Citadel Central Bank Challenge · 2nd Place finish.

---

Disclaimer: This project was developed for educational and competition purposes. It is not financial advice, monetary policy advice, or a production trading system.