Trending News

A notebook-based data science project for predicting short-term market direction from Russian financial news using transformer-based sentiment classification.

Overview

This project fine-tunes a compact RuBERT model on labeled news texts and classifies each article into one of three classes:

• 0 — Increase
• 1 — Stable
• 2 — Decrease

The predictions are then aggregated by date and combined with market data to build a simple trend indicator and backtesting workflow.

Project goals

• clean and normalize financial news text
• train a Russian-language NLP classifier
• evaluate model quality with standard classification metrics
• aggregate news predictions into a daily indicator
• compare signals with Brent market data

Data

The training workflow in eval_model.ipynb uses a labeled Excel dataset:

• data/text_corpus_news.xlsx — training corpus referenced in the notebook
• data/brent_march_2026.xls — market data used for analysis

Note: the Excel corpus is referenced in the notebook but is not included in the current repository snapshot.

Methodology

The training notebook main.ipynb performs the following steps:

1. loads the annotated dataset
2. builds a single target label from Increase, Stable, and Degrease
3. removes noisy text fragments and formatting artifacts
4. splits the data into train / validation / test subsets
5. tokenizes text with cointegrated/rubert-tiny
6. fine-tunes a sequence classification model
7. saves trained weights to models/model_rubert-tiny1

The custom dataset wrapper is CustomDataset, and evaluation is computed in compute_metrics.

Model

• Base model: cointegrated/rubert-tiny
• Framework: PyTorch + Hugging Face Transformers
• Task: multi-class text classification
• Classes: increase / stable / decrease
• Training setup: mixed precision enabled (fp16=True)

Installation

Install dependencies from requirements.txt:

pip install -r requirements.txt

For notebook execution and Excel loading, the following may also be needed:

pip install jupyter openpyxl

How to run

1. Train the model

Open eval_model.ipynb and run the cells in order.

2. Run inference and analysis

Open main.ipynb to test the end-to-end workflow, including:

• run_pipeline
• detect_trend
• merge_with_market
• generate_signals

Example: load saved weights

import torch
from transformers import BertTokenizerFast, BertForSequenceClassification

tokenizer = BertTokenizerFast.from_pretrained("cointegrated/rubert-tiny")
model = BertForSequenceClassification.from_pretrained(
    "cointegrated/rubert-tiny",
    num_labels=3
)

model.load_state_dict(torch.load("models/model_rubert-tiny1", map_location="cpu"))
model.eval()

Evaluation

The project tracks the following metrics during validation:

• Accuracy
• F1-score
• Precision
• Recall

See the metric calculation in compute_metrics and the reported outputs in eval_model.ipynb.

Visualization

An example of calculating the indicator relative to Brent for March 2026:

Limitations

• notebook-centric workflow
• training corpus is not stored in the repository
• compact model chosen for limited resources rather than maximum accuracy
• current pipeline is research-oriented and should be validated further before production use

Future improvements

• move preprocessing and inference into reusable Python modules
• add experiment tracking and configuration files
• improve train/validation/test splitting strategy
• add CLI or API for inference
• document final benchmark results in a separate report

License

Distributed under the MIT License.