michael-zus/fintweet-sentiment-2025

FinTweet Sentiment Dataset 2025

A curated dataset of financial tweets enriched with market data for sentiment-driven trading signal classification.

Dataset Description

This dataset combines social media signals from financial Twitter accounts with real-time market data from Interactive Brokers to create labeled samples for 3-class sentiment classification (BUY/HOLD/SELL).

Dataset Summary

Property	Value
Time Period	2024-2025
Total Samples	43,468
Target Variable	3-class: BUY, HOLD, SELL
Split Strategy	Temporal 70/15/15

Class Distribution

Class	Count	Percentage
BUY	16,666	38.3%
HOLD	10,148	23.3%
SELL	16,654	38.3%

Supported Tasks

• Text Classification: Predict trading signal (BUY/HOLD/SELL) from tweet text and market context
• Multimodal Learning: Combine text embeddings with numerical market features
• Financial Sentiment Analysis: Train domain-specific sentiment models

Dataset Structure

Data Fields

Text Feature:

• text (string): Raw tweet text with ticker mentions ($AAPL, $TSLA, etc.)

Target:

• label_1d_3class (ClassLabel): Trading signal based on 1-day forward return
  • BUY (0): Return > +0.5%
  • HOLD (1): -0.5% ≤ Return ≤ +0.5%
  • SELL (2): Return < -0.5%

Categorical Features:

Feature	Description	Example Values
author	Tweet source account	StockMKTNewz, wallstengine, AIStockSavvy
category	News category	Earnings, M&A, Breaking News, Personnel Changes
session	Market session	premarket, regular, afterhours
market_regime	Market condition	trending, volatile, calm
sector	Stock sector (GICS)	Technology, Healthcare, Financials
market_cap_bucket	Market cap size	mega_cap, large_cap, mid_cap, small_cap

Numerical Features (computed at tweet time, no look-ahead bias):

Feature	Description	Lookback
volatility_7d	7-day price volatility	7 days
relative_volume	Volume vs 20-day average	20 days
rsi_14	Relative Strength Index	14 days
distance_from_ma_20	Distance from 20-day MA	20 days
return_5d	5-day momentum	5 days
return_20d	20-day momentum	20 days
above_ma_20	Binary: price > MA20	20 days
slope_ma_20	MA20 trend direction	20 days
gap_open	Overnight gap	1 day
intraday_range	Day's high-low range	1 day

Metadata:

• timestamp (string): Tweet timestamp (US Eastern Time)
• ticker (string): Stock ticker symbol
• tweet_url (string): Original tweet URL

Data Splits

Split	Samples	Time Period
train	30,428	Earliest 70%
validation	6,520	Middle 15%
test	6,520	Latest 15%

Splits are temporal (chronologically ordered) to prevent data leakage and simulate real-world deployment.

Usage

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("michael-zus/fintweet-sentiment-2025")

# Access splits
train = dataset["train"]
val = dataset["validation"]
test = dataset["test"]

# Example: Get text and labels
texts = train["text"]
labels = train["label_1d_3class"]

# Example: Filter by sector
tech_samples = train.filter(lambda x: x["sector"] == "Technology")

Training with Transformers

from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments
from datasets import load_dataset

# Load dataset
dataset = load_dataset("michael-zus/fintweet-sentiment-2025")

# Load FinBERT tokenizer and model
tokenizer = AutoTokenizer.from_pretrained("ProsusAI/finbert")
model = AutoModelForSequenceClassification.from_pretrained("ProsusAI/finbert", num_labels=3)

# Tokenize
def tokenize(examples):
    return tokenizer(examples["text"], truncation=True, padding="max_length", max_length=128)

tokenized = dataset.map(tokenize, batched=True)

# Train
training_args = TrainingArguments(
    output_dir="./finbert-tweet-classifier",
    num_train_epochs=3,
    per_device_train_batch_size=32,
    evaluation_strategy="epoch",
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized["train"],
    eval_dataset=tokenized["validation"],
)

trainer.train()

Data Collection

Sources

• Tweets: Twitter financial accounts (@StockMKTNewz, @wallstengine, @amitisinvesting, @AIStockSavvy, @fiscal_ai, @EconomyApp)
• Market Data: Interactive Brokers TWS API (S&P 500 + Russell 1000 tickers)

Label Generation

1. Identify entry price (first available bar after tweet)
2. Calculate return to next market open
3. Classify: BUY (>+0.5%), HOLD (±0.5%), SELL (<-0.5%)

Quality Filters

All samples have been validated for:

• Valid entry price at tweet time
• Valid forward price for return calculation
• Market was open (no weekends/holidays)
• No data gaps in price data

Considerations

Biases

• Source accounts may have inherent biases in coverage
• Large-cap stocks are more frequently mentioned
• US market hours only (Eastern Time)

Limitations

• Historical data only (2024-2025 period)
• Twitter-specific language patterns and conventions
• May not generalize to other social platforms or time periods

Citation

@dataset{fintweet_sentiment_2025,
  title={FinTweet Sentiment Dataset 2025},
  author={michael-zus},
  year={2025},
  url={https://huggingface.co/datasets/michael-zus/fintweet-sentiment-2025},
  note={Financial tweets with market data for trading signal classification}
}

License

MIT License - Free to use for research and commercial purposes with attribution.