What this is. A list of the data that quant finance research actually runs on. For each of 4,000 scored papers in the archive (dated 1998 to November 2026) we extracted the datasets the paper says it uses, with the sample period where the paper states one. That gives 7,282 dataset mentions. The tables below count the sources that recur, and the full list at the end is searchable. The names are extracted by a language model from the paper text and are not hand-checked, so read the counts as a careful first pass and read the caveats before quoting them.
The numbers at a glance
| Measure | Value |
|---|---|
| Papers covered | 4,000 |
| Dataset mentions | 7,282 |
| Distinct dataset names | 6,781 |
| Mentions with a stated sample period | 3,572 (49%) |
| Papers naming exactly one dataset | 2,127 (53%) |
| Papers naming three or more | 812 (20%) |
| Most datasets named by one paper | 8 |
| Papers naming a synthetic or simulated dataset | 790 (20%) |
| Papers naming only synthetic or simulated data | 468 (12%) |
By quadrant, 2,270 of the papers are Holy Grail, 839 Street Traders, 724 Lab Rats and 167 Philosophers. The mix leans empirical because a paper with no data has nothing to list.
Three things stand out.
Free data carries the field. Yahoo Finance is named in 317 papers (7.9%), more than CRSP (175) and Bloomberg (108) combined. FRED, the Kenneth French library, Kaggle and exchange APIs add to the same pattern. This matters when you read a result: free daily bars are adjusted after the fact, drop delisted names and carry no point-in-time history, which is the subject of our guides on free versus paid market data and look-ahead bias and point-in-time data.
Names are almost never reused. 6,781 distinct names for 7,282 mentions means nearly every paper describes its data in its own words: “Historical stock price data”, “Dataset 1”, “US equities panel”. Only a handful of names recur verbatim, led by “CRSP daily stock file” (17 mentions) and “LOBSTER” (13). A reader who wants to rebuild a sample often has less to go on than the dataset name suggests. Our guide to finding the datasets behind quant papers covers what to do next.
Crypto venues are now standard data sources. Binance appears in 121 papers, more than any paid equity database except CRSP. Uniswap (55) is ahead of LOBSTER (47) and NYSE TAQ (38). Venue data is free, tick-level and complete for that venue, which explains much of its appeal for microstructure work.
How the counts are made
Each count is the number of papers, out of 4,000, in which a pattern for that source matches. For databases, vendors and venues the pattern is matched against the dataset name and against the sentence from the paper that supports it, because papers usually name the provider in the sentence and not in the label (“daily closing prices … from Yahoo Finance”). For index universes, reference series and benchmark sets the pattern is matched against the dataset name only, since a sentence can mention an index in passing.
Limits to keep in mind:
- Lower bounds. A paper that says “historical price data” without naming a source is counted nowhere.
- Mentions, not licences. A supporting sentence can name a source the paper only compares against. The example paper in each row is the highest-rigor paper that matched, and is not a claim that the source is central to it.
- Overlap. One paper can count toward several rows, for example CRSP, Compustat and WRDS together.
- Extraction errors. Names come from a language model reading the paper. Spot checks look right, and nothing here was verified row by row.
Commercial and academic databases
WRDS is counted only when the paper names it. Most CRSP, Compustat, TAQ, OptionMetrics and I/B/E/S users reach those through WRDS without saying so, so the WRDS row understates its real footprint. The newer developer-first vendors are barely visible: Databento and Polygon.io appear in 8 papers each. If you are choosing a feed, the market data vendor guide compares them on fields and price.
Crypto, DeFi and prediction markets
| Source | Papers | Share | Example paper (highest rigor) |
|---|---|---|---|
| Binance | 121 | 3.0% | Demand Under a Mechanical Price: Elasticity and Welfare in Ethereum… (9) |
| Uniswap | 55 | 1.4% | Measuring Arbitrage Losses and Profitability of AMM Liquidity (8.5) |
| CoinMarketCap | 30 | 0.8% | From Hypotheses to Factors: Constrained LLM Agents in Cryptocurrenc… (8.5) |
| CoinGecko | 22 | 0.6% | ASRI: An Aggregated Systemic Risk Index for Cryptocurrency Markets (8.8) |
| Coinbase | 20 | 0.5% | Does a Trading Policy Need to Know What It Trades? Microstructure I… (9) |
| Polymarket | 19 | 0.5% | The Anatomy of a Decentralized Prediction Market: Microstructure Ev… (9.5) |
| Deribit | 9 | 0.2% | Quantifying Repair Risk in Crypto Option Surfaces (8) |
| Kalshi | 9 | 0.2% | Hard to Fool: A Bounded LLM Research Loop for Quantitative Finance,… (9) |
| Kraken | 8 | 0.2% | Cryptocurrency as an Investable Asset Class: Coming of Age (8.5) |
| OKX | 8 | 0.2% | Deep Learning Models Meet Financial Data Modalities (8) |
| Bybit | 7 | 0.2% | Exploring Microstructural Dynamics in Cryptocurrency Limit Order Bo… (8.5) |
| Hyperliquid | 7 | 0.2% | Autodeleveraging as Online Learning (8.5) |
Index universes and reference series
The S&P 500, as an index series, a constituent universe or through SPX and SPY, is named as a dataset in 484 papers, about one in eight. The Chinese index universes (CSI 300 in 56 papers, CSI 500 in 22) come next among equity universes, driven by the stock-prediction literature.
Text and machine-learning benchmark sets
| Source | Papers | Share | Example paper (highest rigor) |
|---|---|---|---|
| News articles and headlines | 142 | 3.5% | Thinking Fast, Trading Late: System One Decision Models vs the Market (9) |
| Twitter / X and StockTwits | 44 | 1.1% | FinDPO: Financial Sentiment Analysis for Algorithmic Trading throug… (8.5) |
| Earnings-call transcripts | 21 | 0.5% | Corporate Earnings Calls and Analyst Beliefs (8.5) |
| Financial PhraseBank | 11 | 0.3% | FinDPO: Financial Sentiment Analysis for Algorithmic Trading throug… (8.5) |
| StockNet / ACL18 | 10 | 0.2% | Ploutos: Towards interpretable stock movement prediction with finan… (8.5) |
| FNSPID | 7 | 0.2% | FNSPID: A Comprehensive Financial News Dataset in Time Series (8.5) |
| 7 | 0.2% | More Mentions, Less Information? Diagnosing Identity and Alignment… (9) | |
| FI-2010 | 6 | 0.1% | Temporal Kolmogorov-Arnold Networks (T-KAN) for High-Frequency Limi… (9) |
| FiQA | 5 | 0.1% | Instruct-FinGPT: Financial Sentiment Analysis by Instruction Tuning… (8.5) |
| FinQA / ConvFinQA | 4 | 0.1% | VERAFI: Verified Agentic Financial Intelligence through Neurosymbol… (7) |
Fixed benchmark sets are rare. FI-2010 for limit order books, Financial PhraseBank for sentiment and StockNet/ACL18 for movement prediction are the closest things to shared test beds, and each appears in about ten papers or fewer. Most results in the archive cannot be compared on a common dataset.
Synthetic and simulated data
790 papers (20%) name a synthetic or simulated dataset, and for 468 of them (12%) it is the only data named. Simulation is the right tool for a method paper that needs a known ground truth. It is not evidence that a strategy works, which is why the score guide treats a simulation-only study differently from a backtest on market data.
The full list
Every dataset mention, as the paper names it, with the stated sample period and a link to the paper’s page. Search matches the dataset name, the sample period, the paper title and the paper identifier.
Loading the list…
| Dataset as the paper names it | Sample period | Paper | Rigor | Quadrant |
|---|
The same data is served as JSON for scripts and agents: a papers array (id, title, page, year, empirical rigor, quadrant) and a rows array (dataset name, sample period, index into papers).
Questions to ask about a paper’s data
- Is the source named, or only described (“historical price data”)?
- Is the sample period stated? Half the mentions here have none.
- Is the universe point-in-time, with delisted names included, or today’s constituents projected backwards?
- Are prices adjusted, and with which vendor’s adjustment?
- If the data is synthetic, is there any test on market data at all?
- If the data is one crypto venue, do the results depend on that venue’s fees and matching rules?
- Can you obtain the same data, at what cost, and under what licence?
The foundations: finding the datasets behind quant papers → · market data types → · market data vendors → · free versus paid market data → · versioning datasets and backtests → · Run your numbers: tick-data storage sizer → · More data: the scored-paper dataset → · papers with code →
Method: dataset names, sample periods and supporting sentences come from the evidence cards built for each scored paper. Source families are counted by pattern matching, as described above, and the counts are a snapshot taken on 16 November 2026. Browse the flowchart library, the guides, and the blog.