Finance_AI

KO | EN

gitlite — search

#swift #shell #git #java #editor #monitoring #html #chatbot #nextjs #claude-code

(★ 9)

An explainable AI system for analyzing financial reports and generating data-driven investment insights using NLP and machine learning.

File Explorer

ZIP Download

- streamlit_app.py
- 0000.parquet
- report.pdf
- Sentences_AllAgree.txt
- - eval_sentiment.json
- evaluate_sentiment.py
- run_analysis.py
- train_sentiment.py
- - __init__.py
  - __main__.py
  - config.py
  - data.py
  - pipeline.py
.gitignore
README.md
requirements.txt

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

Research tool that reads a U.S. public-company financial PDF and simulates an **equity-research analyst workflow**: section selection, financial metric extraction, feature engineering, weighted impact scoring, and fully traceable explainability.

> **Not financial advice.** The system supports research; it does not make investment decisions for you.

## Design principle

This is **not** “classify sentences and average sentiment.”  
Sentiment is only a supporting signal. The score is driven by **structured financial features** (Revenue Growth, EPS, Cash Flow, Debt, Guidance, Legal Risk, …), each with an explicit weight. Every point on the final 0–10 score is traceable.

## Why this architecture

| Component | Choice | Rationale |
|-----------|--------|-----------|
| Sentiment model | **TF-IDF + Logistic Regression** | Trained on PhraseBank; simple & explainable |
| Feature engine | Rule + metric extraction | Turns “Revenue +12%” into a weighted feature, not just Positive |
| Relevance filter | Embeddings + keywords | Drops boilerplate before scoring |
| Section detection | 10-K/10-Q heading rules | Prefer MD&A, liquidity, risks over TOC/legal filler |
| Scoring | Feature-weighted, traceable points | Analyst-style contribution table from baseline 5.0 |
| UI | **Streamlit** | Upload PDF → score + drivers + evidence |

## Pipeline

```text
PDF → extract → clean → section detection → relevant sections
  → sentences → financial metrics/features → relevance/impact
  → sentiment (support) → weighted feature score → explainability
```

### Feature weights (examples)

| Feature | Weight |
|---------|--------|
| EPS / Earnings | 0.14 |
| Revenue Growth | 0.12 |
| Cash Flow | 0.12 |
| Debt / Leverage | 0.10 |
| Guidance / Outlook | 0.10 |
| Risk Factors | 0.08 |
| … | … |

## Project layout

```text
app/streamlit_app.py          # Web UI
scripts/train_sentiment.py    # Train TF-IDF + logistic regression
scripts/evaluate_sentiment.py # Quick held-out metrics
scripts/run_analysis.py       # CLI analysis
src/investment_ai/            # Core library
  pdf/                        # extract → clean → sentences
  nlp/                        # sentiment (logreg) + relevance
  scoring/                    # aggregation + explainability
  pipeline.py                 # end-to-end orchestration
data/
  Sentences_AllAgree.txt      # Financial PhraseBank (gold labels)
  0000.parquet                # Extra continuous-score sentences
  report.pdf                  # Sample report (optional)
```

## Setup

```bash
python -m pip install -r requirements.txt
```

Train the sentiment model before running analysis (required once):

```bash
python scripts/train_sentiment.py
```

Relevance embeddings still download once from Hugging Face (`sentence-transformers/all-MiniLM-L6-v2`).

## Train / evaluate sentiment

```bash
python scripts/train_sentiment.py
```

- Builds **TF-IDF** features from the **training** split only  
- Trains **multinomial logistic regression** on the labeled training split  
- Saves the checkpoint to `models/sentiment/logreg-scratch/`  
- Reports held-out PhraseBank accuracy / macro-F1 and top weighted features per class

```bash
python scripts/evaluate_sentiment.py
```

```bash
python scripts/train_sentiment.py --phrasebank-only
```

Optional flags: `--C`, `--ngram-max`, `--min-df`, `--max-features`.

## Analyze a PDF (CLI)

```bash
python scripts/run_analysis.py data/report.pdf
python scripts/run_analysis.py path/to/10k.pdf --json
```

## Web UI

```bash
streamlit run app/streamlit_app.py
```

Upload a PDF → receive score, recommendation, positive/negative factors, and the sentences that drove the result.

## Notes on data

- **PhraseBank AllAgree** is the primary gold set (positive / negative / neutral).  
- `0000.parquet` scores are mapped to ternary labels with thresholds (±0.25) and used as optional extra training data.  
- Sentiment uses **TF-IDF + logistic regression**; coefficients are inspectable per class.

Finance_AI

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Finance_AI

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README