Featured Dataset: The Trove Ledger
Collection
Definitive successor to “preliminary v2” – extended, curated, and actively updated.
•
1 item
•
Updated
•
1
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: ValueError
Message: Bad split: daily. Available splits: ['validation']
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 99, in get_rows_or_raise
return get_rows(
^^^^^^^^^
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/src/services/worker/src/worker/utils.py", line 61, in get_rows
ds = load_dataset(
^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/load.py", line 1409, in load_dataset
return builder_instance.as_streaming_dataset(split=split)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1232, in as_streaming_dataset
raise ValueError(f"Bad split: {split}. Available splits: {list(splits_generators)}")
ValueError: Bad split: daily. Available splits: ['validation']Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
🗃️ TroveLedger — Financial Time Series Dataset
A growing ledger of accumulated market history.
Year-End Reflection — how TroveLedger came to be
Over the past year, TroveLedger has gradually taken shape — not as a sudden idea, but as the result of a series of practical needs becoming increasingly clear.
After spending more time engaging seriously with financial markets, the idea emerged around the turn of the previous year to experiment with training custom AI trading models. Very quickly, one fundamental requirement became apparent: reliable, well-structured historical market data.
Finding such data turned out to be the first real challenge.
Existing datasets on Hugging Face were explored early on, but it became clear that many of them did not meet the requirements needed for long-term model training — whether due to gaps, inconsistent formats, limited history, or missing intraday resolution. Paid datasets were considered, but ultimately the decision was made to experiment with collecting data independently using yfinance.
Some of the earliest datasets — particularly for commodities — went through multiple transformations and revisions. A few of them have not yet found their way into the current TroveLedger dataset, though they may still be integrated over time.
By the end of May, a stable and consistent data format based on Parquet files had been established. From that point onward, the first stocks began accumulating continuously, gap-free, and in a uniform structure. What started as a small list of equities quickly grew to several hundred within a matter of weeks.
At that stage, publishing the data as a public Hugging Face dataset became a natural step:
- the data contained nothing private or proprietary,
- it allowed easy access from different locations,
- and it made collaboration possible, as the project had grown beyond a strictly solo effort.
The dataset was initially released under a preliminary status. Somewhat unexpectedly to me, it began to attract a steady stream of external downloads, indicating genuine interest. After refinements to the data pipeline, this evolved into Preliminary v2, accompanied by a promise that this transitional phase would eventually be replaced by a stable, long-term dataset.
That transition, however, took longer than anticipated. While the pipeline itself matured significantly during this time, the more substantial challenge turned out to be organizational rather than technical: curating which stocks should be included.
One thing was always clear — and remains so today: TroveLedger does not aim to include “all stocks”.
It is not a comprehensive finance database and not a competitor to yfinance. Instead, the goal is to provide long-accumulated minute-, hourly-, and daily-level time series for interesting, liquid, and relevant assets — data that is genuinely suitable for training AI models. Illiquid penny stocks, for example, offer little value in this context.
This naturally led to the realization that such assets are already professionally curated elsewhere: in indices.
The missing step, therefore, was extending the data pipeline to work with index component lists — collecting them, maintaining them, and systematically integrating their constituents. Once this was implemented, the size of the dataset grew rapidly. Adding everything at once proved impractical, both technically and logistically, which led to the decision to introduce new assets gradually, index by index.
The first non-preliminary release of the dataset finally went live on December 17, 2025. By that point, Preliminary v2 had already reached a stable level of well over 1,500 downloads per month.
The current approach builds on that foundation: new stocks are added regularly in structured batches, each corresponding to a complete index and accompanied by a simple, readable component list.
As the project gained more weight and consistency, it also gained a name: TroveLedger.
Along with it came the idea of visually representing the project through a goblin figure — inspired by the meticulous vault-keepers of Gringotts — who serves as the fictional keeper of the ledger itself. Each newly added index is now accompanied by its own image, a small creative detail that reflects the personal nature of the project and helps give it a distinct identity.
With the turn of the year, the ledger turns a page
TroveLedger has since continued to grow through regular updates and the steady addition of further indices. With the turn of the year, a new phase begins. Over the past days, download numbers have increased noticeably and are on track to surpass those of Preliminary v2.
The ledger remains open.
🔔 Latest Dataset Update
Date: 2026-01-02 New addition: 🇦🇺 ASX 200 (Australian Securities Exchange, Australia)
Today, the dataset expands with the ASX 200 — Australia’s primary equity benchmark and the central reference point of its capital markets.
Composed of the 200 largest and most liquid companies listed in Australia, the index reflects an economy shaped by resources, finance, and global trade exposure.📊 Market context:
Region: Oceania — Australia
Scope: Large-cap benchmark (ASX 200)
Sector exposure: Financials, Materials, Energy, Healthcare, Industrials
Data coverage: Minute, hourly, and daily OHLC data📈 What this means:
A clear view into a resource-driven, globally linked equity market with distinct sector concentration.🔜 What’s next:
Continued expansion of regional coverage and structural market diversity.
Click to expand
- More information about previous additions can be found further below in The Growing Treasury
Recent Index Additions
Date Index Region Symbols 2026-01-02 ASX 200 Australia 🇦🇺 200 2025-12-30 OMX Stockholm 30 Sweden 🇸🇪 30 2025-12-29 TSX (S&P/TSX Composite) Canada 🇨🇦 222 2025-12-24 SMI 🇨🇭 Switzerland 20 2025-12-23 NIFTY 50 🇮🇳 India 50 2025-12-22 FTSE 100 🇬🇧 United Kingdom 100 2025-12-19 S&P 500 🇺🇸 US 503 2025-12-18 Hang Seng Index 🇭🇰 Asia 82 2025-12-17 EURO STOXX 50 🇪🇺 Europe 50
📌 Overview
TroveLedger is a public financial time series dataset focused on long-term accumulation of high-quality intraday data.
The dataset provides OHLC and volume data at multiple time resolutions and is designed primarily for machine learning, quantitative research, and systematic trading experiments.
Unlike many freely available data sources, TroveLedger emphasizes continuity over time, especially for minute-level data.
Scale & Granularity
- Total: Over 40 million rows across all symbols and resolutions (growing rapidly)
- Per symbol: Varies significantly – from <1,000 rows (young stocks, daily) to >500,000 rows (established stocks, minute-resolution)
- Ideal for both focused single-symbol training and large-scale multi-market models
🔑 What makes TroveLedger different
High-resolution intraday data is difficult to obtain from free sources over extended periods.
Typical public data access (e.g. via yfinance) provides:
- Daily candles: often spanning decades
- Hourly candles: roughly one year into the past
- Minute candles: usually limited to the most recent 7 days
Repeatedly downloading rolling 7-day windows results in short, fragmented histories that are poorly suited for training models on intraday behavior.
TroveLedger takes a different approach:
- Minute-level data is accumulated continuously
- Time series are extended, not replaced
- Over time, this results in months of gap-free minute data per instrument
This accumulated depth forms a substantially more reliable foundation for intraday research and model training.
🧱 Data Integrity Philosophy
TroveLedger prioritizes continuity over frequency.
The primary goal is not to fetch data as often as possible, but to ensure that once a time series starts, it remains gap-free.Minute-level data is accumulated incrementally over time, creating long, uninterrupted histories that are not obtainable from fresh API queries alone.
This makes the dataset particularly suitable for model training, backtesting, and regime analysis.
📦 Dataset Structure
The dataset is organized as follows:
- /data/{category}/{symbol}/{symbol}.{interval}.valid.parquet
Where:
{category}: e.g.,equities/us,indices/sp500,indices/eurostoxx50(growing with new indices){symbol}: Stock ticker (e.g., AAPL, BMW.DE){interval}: One ofdays(daily),hours(hourly), orminutes(1-minute)
The .valid suffix indicates that these files have passed quality checks and are ready for use. Only these cleaned, validated files are included in the dataset – temporary or intermediate files from the pipeline are excluded.
Tip for users: The .valid part is intentionally kept as a flexible "state" marker. You can easily rename or copy files to add your own states (e.g., .train.parquet or .test.parquet) for train/validation/test splits in your ML workflows. This pattern makes it simple to organize experiments without changing the core data.
Data Instances
Here's an example row from a typical daily Parquet file (e.g., for AAPL.days.valid.parquet):
| symbol | time | open | high | low | close | volume |
|---|---|---|---|---|---|---|
| AAPL | 1704067200 | 192.28 | 192.69 | 191.73 | 192.53 | 42672100 |
timeis a Unix timestamp (e.g., 1704067200 = January 1, 2024, 00:00 UTC).- All prices are in the symbol's native currency (e.g., USD for US equities).
Dataset Creation
Curation Rationale
TroveLedger was created to provide a reliable, expanding source of historical OHLCV data for AI-driven trading research, addressing gaps in continuity and international coverage.
Source Data
All data is sourced from Yahoo Finance via the yfinance Python library. Index components are automatically extracted from Wikipedia pages using a custom API-based pipeline for sustainability.
Data Collection and Processing
- Symbols are selected from major indices (e.g., S&P 500, EURO STOXX 50) and equities.
- Data is fetched at daily, hourly, and 1-minute resolutions, validated for completeness, and stored in Parquet format for efficiency.
- Quality checks remove gaps or anomalies; only ".valid" files are included.
- Updates occur periodically to extend histories and add new indices based on community input.
Who are the source data producers?
Yahoo Finance (public market data). No personal data is included.
🔄 Update Philosophy
The primary objective is data continuity, not guaranteed daily updates.
In particular:
- Daily updates are not guaranteed
- Preventing gaps in accumulated minute data has priority
- Updates are performed on trading days whenever possible
Minute data is updated most frequently to ensure continuity.
Hourly and daily data are updated on a rotation basis to reduce unnecessary repeated downloads and to remain considerate of public data sources. These datasets are guaranteed to be no older than one week.
For most training scenarios, this is fully sufficient. When models are deployed in real-world environments, current market data is typically provided directly by the target trading platform.
📈 Scope & Growth
TroveLedger started with a curated universe of approximately 500 equities inherited from earlier Preliminary datasets.
Going forward:
- Entire indices are added step by step
- The covered universe will grow continuously
- Expansion is performed incrementally to ensure data integrity and operational stability
This gradual approach allows issues to be detected early and handled without disrupting existing data.
🎯 Intended Uses
- Primary Use: Training and evaluating machine learning models for trading strategies and autonomous AI bots.
- Other Uses: Time series analysis, financial research, educational projects, and community-driven extensions.
TroveLedger is suitable for:
- machine learning on financial time series
- intraday and swing trading research
- feature engineering on OHLC data
- backtesting strategies requiring dense intraday history
- exploratory quantitative analysis
⚠️ Limitations & Notes on Data Sources
- Data Freshness: Data is typically a few days old, not real-time.
- Coverage: Not all symbols may have complete historical data, especially for minute-resolution or newly added indices.
- Growth Phase: The dataset is actively expanding; check for updates on new indices and symbols.
- Not financial advice: This dataset is for research and educational purposes only. Past performance is no guarantee of future results.
Data is derived from publicly accessible market data sources (e.g. via yfinance).
While care is taken to ensure consistency and continuity, this dataset is provided as-is and without guarantees regarding completeness or correctness.
Users are responsible for verifying suitability for their specific use cases and for complying with the terms of the original data providers.
📜 License & Usage
This dataset is provided solely for non-commercial research and educational purposes.
The data is retrieved from public sources via the yfinance library (Yahoo Finance). All rights remain with the original data providers.
Redistribution of this dataset is not permitted without explicit permission from the original sources.
See the LICENSE file for full details.
NO WARRANTY IS PROVIDED. Use at your own risk.
💬 Feedback, Suggestions & Community Support
TroveLedger is a growing, community-driven project providing high-quality OHLCV data for training AI models on financial markets and trading strategies. Your input makes it better!
- What are you building? I'd love to hear how you're using TroveLedger! Share your projects, trading bot ideas, ML models, or research directions – it motivates me to keep expanding and might inspire others.
- Desired indices: Which major indices are you waiting for most? I'll prioritize based on demand and feasibility.
- Helping expand indices: The pipeline uses the Wikipedia API to automatically extract components. It works best with a structured table containing both company names and clean, yfinance-compatible ticker symbols.
- Simply share the Wikipedia page URL (any language) for your desired index.
- If the table needs tweaks (e.g., missing or unclear ticker column, prefixes in symbols), improving it on Wikipedia is the most sustainable way – the global community then keeps it updated long-term!
- Once ready, post the link here, and I'll integrate it quickly.
Interested in a deeper dive into the exact table format and config options my pipeline supports (with examples like zero-padding, suffixes, or language overrides)? Let me know – if there's demand, I'll create a dedicated guide soon!
Join the discussion in Hugging Face Discussions.
🏛️ The Growing Treasury
Watch TroveLedger expand across global markets – a visual chronicle of added indices:
🇦🇺 ASX 200 (January 2, 2026) – Weighed by Earth and Capital
Across vast distances and resource-rich ground, the ASX 200 captures an equity market anchored in tangible assets and institutional capital. Mining conglomerates, major banks, energy firms, and healthcare leaders dominate the index, forming a market profile distinct from technology-heavy regions.
This addition extends TroveLedger’s reach into Oceania, preserving a market where commodities, yield, and global demand cycles leave clear historical traces across intraday and long-term data.
🇸🇪 OMX (December 30, 2025) – Order in the Nordic Ledger
In the measured calm of Northern Europe, the Stockholm Stock Exchange (OMX) records value through discipline, transparency, and long-term orientation. Industrial groups, financial institutions, and globally oriented consumer firms dominate the OMX Stockholm 30, forming a compact yet internationally relevant market profile.
This entry adds a distinctly Nordic balance to TroveLedger — one shaped by export strength, institutional stability, and methodical capital allocation across intraday and long-horizon views.
🇨🇦 TSX (December 29, 2025) – Beneath the Surface of Canadian Capital
Deep underground, where resources are extracted and value is carefully recorded, the Toronto Stock Exchange (TSX) reflects the structural foundations of the Canadian economy. Banks, miners, energy producers, and industrial firms form the backbone of the S&P/TSX Composite, making it a distinctive counterweight to more tech-heavy global indices.
This entry extends TroveLedger’s North American coverage beyond the United States, adding a market shaped by commodities, capital discipline, and long-cycle industries — all captured across consistent intraday and long-horizon timeframes.
🇨🇭 SMI (December 24, 2025) – Alpine quality meets market stability
The Swiss Market Index (SMI) has been added to TroveLedger, bringing the premier blue-chip index of Switzerland into our global dataset.
Representing 20 of the largest and most liquid companies listed on the SIX Swiss Exchange — including giants like Nestlé, Roche, and Novartis — the SMI offers a unique exposure to one of the world’s most stable and innovation-driven economies.
The SMI reflects Switzerland’s enduring role as a benchmark for quality, resilience, and long-term value.
🇮🇳 NIFTY 50 (December 23, 2025) – India takes center stage
The NIFTY 50 Index from India has been incorporated into TroveLedger, enriching the dataset with one of South Asia’s most referenced equity benchmarks. It represents 50 of the largest and most liquid Indian stocks listed on the National Stock Exchange.
🇬🇧 FTSE 100 (December 22, 2025) – Britain weathers the storm
The FTSE 100 represents 100 of the most capitalized and liquid firms on the London Stock Exchange, spanning finance, energy, consumer goods, healthcare, and industrial sectors.
As the UK is no longer part of the European Union, this addition extends TroveLedger’s European coverage beyond the Eurozone without overlap with previously added indices.
🇺🇸 S&P 500 (December 19, 2025) – America answers the call
The complete S&P 500 Index (503 constituents) has been fully integrated, adding 173 new symbols.
This provides the premier US large-cap benchmark with extended intraday histories – ideal for multi-sector trading bot training.
🇭🇰 Hang Seng Index (December 18, 2025) – Asia opens its doors
The Hang Seng Index (HSI) adds 82 entirely new symbols – major Hong Kong-listed companies with strong China exposure across finance, tech, energy, and consumer sectors.
🇪🇺 EURO STOXX 50 (December 17, 2025) – Europe uncovers its treasures
The EURO STOXX 50 introduces 50 blue-chip companies from the Eurozone, spanning multiple countries and sectors – a cornerstone for European market exposure.
🔖 Citation
If you use TroveLedger in your work, please cite it as:
@dataset{Traders-Lab_TroveLedger_2025,
author = {Traders-Lab},
title = {TroveLedger Financial Time Series Dataset},
year = {2025},
url = {https://huggingface.co/datasets/Traders-Lab/TroveLedger}
}
Support the Treasury Expansion 🚀
If you're training AI models or building quant pipelines with the TroveLedger dataset, consider supporting further global indices:
Thank you for helping grow the trove! 💰🗝️
🔚 Final note
TroveLedger grows by geography, not by noise — each market a layer in the record.
- Downloads last month
- 1,251