The 20 Best Financial Datasets, Scored on the Three Defects That Decide If You Can Use Them

Every data team we work with has run the same experiment. Somebody pulls ten years of price history, joins it to fundamentals, tests an idea, and the result comes back better than it has any right to be. The instinct is to celebrate. In our experience the dataset is wrong far more often than the idea is right.

There are three things worth checking before you trust any financial dataset, and all three are silent. Nothing in the file tells you they are there. The first is whether companies that failed are still in it. The second is whether the numbers are what was published at the time or what got corrected later. The third is what the licence actually lets you do with the data once you have it. We have watched all three cost teams a quarter, and the third is the only one you cannot engineer your way out of afterwards.

So this list is organised around those three questions rather than around price or rate limits. Twenty datasets, grouped by the kind of data they hold, each one scored on all three defects. Where we could not establish an answer, the entry says so instead of guessing.

What counts as a financial dataset, and the four kinds you will actually meet

A financial dataset is any structured record of economic or market activity you can load and compute on. That definition is broad enough to be useless on its own, which is why the term collects such a strange mix of things: a central bank time series, a tick-level order book, a folder of quarterly filings, and a CSV of scraped headlines all answer to it.

In practice there are four classes, and they fail in different ways. Sorting a dataset into the right class before you evaluate it saves most of the wasted effort in this category, because the questions worth asking are class-specific.

Most disappointments in this category come from evaluating a dataset against the wrong question. A team benchmarks two price datasets on how many tickers each returns, picks the larger one, and never asks whether either includes the companies that went bankrupt. Both numbers were real. Neither answered the question that mattered.

The three defects that make a financial dataset unusable

These are not edge cases and they are not exotic. They are ordinary properties of ordinary datasets, and each one is silent: nothing in the file tells you it is there. Every entry in this list is scored against all three, and where the answer could not be established, the entry says that rather than guessing.

Defect 1. Survivorship bias

A dataset has survivorship bias when it only contains securities that still exist. Build a universe from today's listed tickers, pull ten years of history for each, and you have quietly excluded every company that went bankrupt, got acquired, or was delisted during those ten years. What remains is a sample of survivors, and a strategy tested on survivors will look better than it is.

The size of the effect is easy to underestimate until it is counted. CRSP maintains a database it explicitly markets as survivor-bias-free, covering more than 64,000 US open-end mutual funds. Over 31,000 of those are delisted. Roughly half the population of that dataset consists of funds that no longer trade, and every one of them is missing from a naive sample.

Defect 2. Look-ahead bias, and the as-filed problem

Look-ahead bias is any situation where a model sees information it could not have had at the time. In financial data it arrives through two doors, and both are easy to walk through without noticing.

The first door is the date you join on. A company's fiscal quarter ends on 31 March and the 10-Q lands several weeks later. Join fundamentals to prices on the period-end date and your model is trading on numbers that were not public for another month or more. The fix is to join on the filing date, which means the dataset has to carry one.

The second door is restatement. Most fundamentals databases hold the current version of each figure, which is to say the version after any subsequent corrections. Testing against those is testing against numbers nobody had at the time. An as-filed or point-in-time dataset preserves what was actually published on the day it was published, which is why point-in-time products exist as separate, more expensive offerings rather than as a default setting.

# Look-ahead: uses figures that were not public on period_end
df = prices.merge(fundamentals, left_on="date", right_on="period_end")

# Correct: join on the date the filing actually became public
df = prices.merge(fundamentals, left_on="date", right_on="filing_date")

Defect 3. Licence scope, which is not the same as price

Free to download and free to use are different claims, and the gap between them is where commercial projects die. A dataset can be freely accessible, freely queryable, and still prohibited from redistribution, from being cached inside your product, or from being shown to your customers.

FRED is the instructive case, because it is both the best free dataset in this category and more restricted than its reputation suggests. Its API terms require a specific notice on any application built with it: "This product uses the FRED® API but is not endorsed or certified by the Federal Reserve Bank of St. Louis." More importantly, FRED aggregates from 126 sources, and some of those individual series are copyrighted by their original owners. The terms flag copyrighted series in their notes, direct users to seek permission from the data owner for anything beyond personal use, and state plainly that you are solely responsible for complying with restrictions the owners impose.

How we picked these datasets and how to read each entry

No vendor paid for placement and none were contacted before publication. The 20 entries are the datasets that recur across this search and that we could confirm are still obtainable in September 2026. Two well-known names failed that second test and are handled in their own section rather than listed as though nothing had changed.

Every entry carries two tables. The first is the specification. The second is the limits table, which prints a verdict on each of the three defects plus the known gaps and how confident we are in the assessment. Ratings deliberately do not appear anywhere on this page, because datasets do not carry meaningful public review corpora and a star rating borrowed from a vendor's software listing would tell you nothing about the data inside it.

What we could not verify. Licence terms for the institutional datasets in Band C sit behind sales contact and are not published, so those rows read "institutional licence, terms not public" rather than carrying a guess. Coverage figures for commercial vendors are vendor-published and dated, not independently audited. Where a defect verdict depends on which product tier you buy, the entry says so instead of collapsing it into a single answer.

The 20 financial datasets at a glance

Find your class in the second column, then read only that band. Datasets in different bands are not substitutes for each other, and the most common shortlisting error in this category is comparing across them.

The 20 best financial datasets, by class

Each entry carries a specification table, a limits table with a verdict on all three defects, and three paragraphs covering what it is, who it fits, and what it cannot do. The third of those is the one worth reading twice.

Band A. Official public datasets

These are published by central banks, regulators, and multilateral institutions, and they are the most underrated datasets in finance. They are free, stable, documented, and in two cases they solve a defect that expensive commercial products charge specifically to fix. The trade is that none of them will price a security.

1. FRED, Federal Reserve Economic Data

What it is. FRED is the St. Louis Fed's aggregation of economic time series from 126 statistical agencies and central banks, wrapped in an API that has been stable for over a decade. The feature that separates it from every other free macro source is vintage preservation: through its ALFRED interface it stores the original release of a revised series, so a backtest can use the unemployment number as it was first printed rather than as it reads after two revisions.

Best for. Any model with a macro input, any regime-detection work, and any project that needs the value as it was known on the date. It is also the correct first stop before paying for macro data, because the overlap with commercial macro products is larger than those products advertise.

What it cannot do. It will not give you a stock price, a company identifier, or a filing. The licence is also less permissive than its reputation: the attribution notice is mandatory, and because FRED aggregates from other bodies, some series carry third-party copyright with the compliance burden explicitly placed on you. Read the notes field before you redistribute a series, not after.

2. SEC EDGAR and the Financial Statement Data Sets

What it is. EDGAR is the SEC's public filing archive, and the Financial Statement Data Sets are quarterly extracts that flatten the XBRL tags from those filings into numeric tables you can join. It is the only free source of as-filed US fundamentals, which makes it the cheapest available fix for the second defect in this article.

Best for. Anyone building a fundamentals backtest who cannot justify an institutional point-in-time licence, and anyone who needs the document rather than the extracted number. Read the access rules before you write the crawler: the SEC publishes a ceiling of 10 requests per second, requires a declared User-Agent identifying you, and states that it reserves the right to limit request rates to preserve fair access.

import requests

# Requests without a declared User-Agent get blocked.
HEADERS = {"User-Agent": "Your Company Name contact@yourdomain.com"}
url = "https://data.sec.gov/api/xbrl/companyconcept/CIK0000320193/us-gaap/Revenues.json"

r = requests.get(url, headers=HEADERS, timeout=30)
r.raise_for_status()

What it cannot do. It covers US issuers and nothing else, and the structured layer only goes back to the start of XBRL mandates, so pre-2009 work means parsing documents. The bigger cost is normalisation. Filers tag the same concept differently, extensions proliferate, and the gap between a raw EDGAR extract and a clean comparable panel is measured in engineering months. That gap is what commercial fundamentals vendors charge for, and it is the same class of work Forage AI's Intelligent Document Processing does on financial filings, at 95% table detection accuracy across table types including non-grid structures.

3. World Bank Open Data

What it is. The World Bank's open data platform, best known for the World Development Indicators, is the widest free cross-country panel available. Its real value is emerging-market depth, where commercial macro vendors are thin and expensive.

Best for. Country risk work, cross-sectional macro research, and any analysis where the unit is a country rather than a security. It pairs naturally with FRED, which is deeper on developed markets and higher frequency.

What it cannot do. Annual frequency rules out most trading applications, and publication lags of a year or more are routine for some indicators. The absence of retained vintages is the important limitation, because indicators here are revised in place, so any historical analysis is quietly using numbers that were not available at the time.

4. IMF Data

What it is. The IMF publishes the harmonised external-sector statistics that make countries comparable to each other: balance of payments, reserves, trade direction, and government finance on consistent definitions. Harmonisation is the product, and it is the reason to use the IMF rather than stitching together national statistics offices.

Best for. Currency and sovereign work, cross-border flow analysis, and any model that needs comparable fiscal or external accounts across many countries. The archived World Economic Outlook vintages are an underused asset, because they let you test what consensus expected at a past date rather than what turned out to be true.

What it cannot do. Reporting lags differ enormously between members, so a panel that looks complete for OECD economies will be sparse elsewhere. The SDMX-based API is also markedly harder to work with than FRED's, and most teams end up wrapping it once and caching the result.

5. ECB Data Portal and Eurostat

What it is. Two open sources treated as one entry because in practice you use them together: the ECB Data Portal for monetary and market series, Eurostat for harmonised real-economy statistics. The euro reference exchange rates published daily by the ECB are the closest thing to an official FX fixing available for free.

Best for. Anything euro-denominated, European rates and credit work, and FX analysis that needs a defensible daily reference rather than a broker quote. It is the European counterpart to FRED, and a macro model covering both regions generally needs both.

What it cannot do. Coverage stops at the EU. The practical obstacle is discovery rather than access: series keys are long, hierarchical, and poorly searchable, so the first day with these APIs is usually spent finding the identifier rather than pulling the data.

Band B. Market price and reference datasets

This is where survivorship bias lives, and where the gap between the free and paid tiers is widest. Every dataset here will hand you a price history. The question that separates them is whether the universe they built that history from includes the companies that stopped existing.

6. CRSP

What it is. CRSP is the academic standard for US security price history, and the reason it holds that position is the one thing it refuses to do: drop the dead. Delisted securities and their delisting returns stay in the files, which is why published finance research is built on it and why a strategy that works on CRSP and fails on a free dataset is usually the free dataset's fault.

Best for. Academic research, and any fund that can justify an institutional subscription and needs a defensible historical result. If you are publishing a number that someone will challenge, this is the source that ends the argument.

What it cannot do. It is not real time, it is US only, and you almost certainly cannot buy it as an individual. Redistribution is prohibited under the subscribing institution's agreement, so nothing derived from it ships inside a commercial product without a separate conversation.

7. Norgate Data

What it is. Norgate sells the thing CRSP is famous for, to people who cannot buy CRSP. Survivorship-bias-free end-of-day history for US, Australian and Canadian equities, with historical index constituency, on an individual subscription. It integrates with the Zipline backtesting stack through a published package, which is how most people encounter it.

Best for. Anyone running long-horizon equity backtests without institutional access. If you have read the survivorship section above and concluded you have a problem, this is the cheapest honest fix in the roster, and it is the entry we were most surprised to find missing from every ranking article on this search.

What it cannot do. It is end-of-day only, and the vendor says so directly: no live quotes, no delayed quotes, no intraday, no tick. Coverage stops at three markets, so a European or Asian universe rules it out. The delivery model is a desktop updater plus plugins rather than a REST API, which is a real integration cost if your stack is cloud-native.

8. Nasdaq Data Link

What it is. Nasdaq Data Link is a marketplace rather than a dataset: many publishers, one API, one authentication model, one billing relationship. The convenience is real and the naming is confusing, because the platform is what Quandl became and a great deal of published code still refers to the old brand.

Best for. Teams that want several commercial datasets without negotiating and integrating each one separately. The single API and consistent schema genuinely reduce integration work, which is worth something when you are combining four vendors.

What it cannot do. It cannot give you a single answer about quality, because there is no such thing as Nasdaq Data Link data; there are only the publishers behind each product. The larger trap is expectation: people arrive looking for the free Quandl catalogue they remember, and most of it is not there any more.

9. Polygon.io

What it is. Polygon serves consolidated US market data at whatever granularity you are willing to pay for, from daily bars to individual trades and quotes. The free tier is the most generous in this band at roughly five requests a minute, which is enough to build and test a pipeline properly before committing.

Best for. Anyone who needs intraday or tick-level US data, options chains, or a real-time feed with a clear upgrade path from a prototype. It is the sensible default for US-only systematic work that has outgrown daily bars.

What it cannot do. Coverage outside the US is limited, and deep history at tick granularity gets expensive quickly. Survivorship is your responsibility here, not the vendor's: delisted tickers exist in the reference data, but a universe assembled from currently-active symbols will silently exclude them.

10. Tiingo

What it is. Tiingo is a clean end-of-day US price dataset with careful corporate-action handling, aimed explicitly at quantitative researchers rather than at traders. Adjustment quality is the differentiator, and it matters more than most people expect, because badly handled splits and dividends produce phantom returns that look exactly like alpha.

Best for. Daily-frequency backtesting where you want reliable adjusted history without an institutional budget. It is the natural step up from a free scraped source for anyone who has started to care whether their price series is actually correct.

What it cannot do. It is not an intraday or low-latency product and its non-US coverage is limited. Test the delisted universe yourself before trusting it for long-horizon backtests, because survivorship completeness is not something the product leads on.

11. EODHD

What it is. EODHD trades US depth for global breadth, covering exchanges that the US-centric providers in this band do not reach at all. It also sells delisted tickers as an explicit product feature, which is rare enough at this price point to be the main reason to shortlist it.

Best for. Anything non-US or multi-region, and any backtest that needs a global universe with dead names included. If your research question spans more than one continent, this band narrows to EODHD fast.

What it cannot do. Data quality is not uniform across its exchange list, and the further you get from major markets the more validation you should budget for. The free tier is too small to evaluate anything seriously, so plan on paying before you know whether it works for you.

12. Alpha Vantage

What it is. Alpha Vantage is the simplest API in this band and the one most tutorials reach for, because it needs a key, a URL, and nothing else. Built-in technical indicators are a genuine convenience if you would otherwise compute them yourself.

Best for. Teaching, demos, and small prototypes where a handful of symbols is the whole requirement. It is the fastest way to get a working example running, which is a real category of usefulness.

What it cannot do. The free tier is unworkable for anything with a universe: 25 calls a day means a mid-cap index takes the better part of a month to fetch once. Survivorship bias is present and undocumented, so it should not be used for long-horizon backtesting at any tier.

13. Yahoo Finance via yfinance

What it is. This is the most used and least defensible entry on the page. There has been no official Yahoo Finance API since 2017, so the libraries everyone reaches for are calling the same undocumented endpoints the website uses, without a licence, a contract, or a guarantee that any of it works tomorrow.

Best for. Personal exploration, teaching, and quick checks where nothing depends on the answer. The breadth is genuinely useful and the convenience is unmatched, which is exactly why it keeps ending up in places it should not be.

What it cannot do. It cannot be a production dependency, and it cannot support a commercial product, because there is no licence granting you the right to use it that way. Survivorship bias here is severe rather than partial, since delisted symbols stop resolving entirely. If a result matters, reproduce it on a licensed source before you act on it. For a free alternative that at least has a stable public interface, Stooq offers decades of end-of-day history without an API key.

Band C. Fundamentals, filings and estimates

Prices tell you what happened. Fundamentals tell you what the company said, and that is where the second defect lives. This is where look-ahead bias is decided, and where the price gap between free and institutional is largest. The expensive products in this band are expensive for one specific reason: they preserve what was known on the day, and normalising thousands of filers into comparable fields is genuinely hard work.

14. Compustat

What it is. Compustat is the standardised fundamentals database that most published finance research runs on, usually joined to CRSP prices. Standardisation is the product: thousands of filers reporting the same concept differently are mapped into consistent fields you can compare across companies and decades.

Best for. Institutional research where the result has to survive scrutiny, and any strategy where the fundamentals need to be as-filed rather than as-corrected. Buying the point-in-time variant is the whole point, because the standard files carry the same restatement problem as everything else in this band.

What it cannot do. It is not accessible without an institutional relationship and it cannot be redistributed. Standardisation also cuts both ways: the vendor's mapping decisions are opinions, and where you need exactly what the filer said, EDGAR is the more faithful source.

15. LSEG Data and Analytics

What it is. LSEG's data business, formerly Refinitiv, is one of the few places to get global cross-asset coverage under a single vendor relationship. Breadth is what you are buying, along with the analyst estimates and reference data that are hard to assemble any other way.

Best for. Institutions that need one vendor to cover many asset classes and regions, and any workflow that depends on estimates or fixed-income reference data. If your data problem spans equities, credit, and FX at once, this band shortens to two or three names fast.

What it cannot do. Nothing here is accessible without a negotiated contract, and the entitlement model means what you are allowed to do with the data is a separate and often harder question than what you receive. Budget for the derived-data and redistribution conversation early, because discovering the limits after you have built the product is the expensive path.

16. Financial Modeling Prep

What it is. Financial Modeling Prep sits between the free scraped sources and the institutional databases, offering statements and prices through one developer-friendly API. Having fundamentals and prices behind the same key removes a whole integration, which is a real saving when the alternative is joining two vendors on inconsistent identifiers.

Best for. Application builders, screeners, dashboards, and analysts who need workable fundamentals without an institutional contract. The 250-call free tier is genuinely enough to build something, which separates it from most of Band B.

What it cannot do. It is not point-in-time, so it will not fix look-ahead bias, and coverage quality thins outside large-cap US names. Spot-check anything load-bearing against the original filing, since independent reviewers report occasional divergence from EDGAR.

Band D. Alternative and unstructured datasets

This band is deliberately short. Alternative data is a large enough subject to deserve its own treatment, and we have given it one in our comparison of top alternative data vendors. What follows is the three sources you can access today without a vendor negotiation, and the licence question that governs all of them.

17. Kaggle financial datasets

What it is. Kaggle hosts an enormous collection of community-uploaded financial datasets, and it is the fastest way in existence to get a modelling idea in front of real numbers. For learning and for competitions it is excellent, and nothing below is an argument against using it for that.

Best for. Prototyping, teaching, benchmarking a technique, and finding out whether an idea is worth pursuing before you pay for data. Treat it as a laboratory, not a supply chain.

What it cannot do. It cannot support a production system or a defensible result, because provenance is usually unstated and licence claims are made by uploaders who often had no right to make them. The specific failure mode is silent: a dataset with undocumented survivorship bias produces a beautiful backtest and no warning, and because the methodology was never published, you cannot audit it after the fact.

18. GDELT

What it is. GDELT monitors global news in many languages and publishes coded events, extracted entities and themes, and tone scores, updated every 15 minutes. It is the largest free news dataset that finance teams actually use, and the event history reaching back to 1979 makes long-horizon work possible.

Best for. Macro and geopolitical signal construction, media-attention measures, and event studies where you need coverage across languages and regions. The BigQuery tables remove most of the infrastructure problem, which is otherwise the reason teams abandon it.

What it cannot do. It will not tell you which company a story is about in a form you can join to a ticker. Entity resolution is the entire cost of using GDELT in finance, and it is the reason commercial news-sentiment vendors exist despite this being free. Resolving 'ABC Corp', 'ABC Corporation' and 'ABC Co Ltd' to one canonical entity is the specific job Forage AI's Entity Matching Agent does, and it is the step between this dataset and anything you can join to a ticker.

19. Google Trends

What it is. Google Trends reports relative search interest over time, and it has a long track record as an attention proxy in published finance research. It is free, immediate, and global, which is a rare combination in this band.

Best for. Retail attention measures, demand nowcasting, and any hypothesis where public interest is plausibly a leading indicator. It works best as one feature among many, not as a signal on its own.

a = trends(kw="inflation", start="2024-01-01", end="2024-06-30")
b = trends(kw="inflation", start="2020-01-01", end="2026-09-01")

a.loc["2024-03-01"], b.loc["2024-03-01"]   # not equal

What it cannot do. It never gives you an absolute search volume, only an index scaled to the window you requested. That normalisation is the trap: pull the same date inside two different date ranges and you get two different numbers, so a series stitched together from overlapping requests is not a real time series unless you rescale it deliberately.

Band E. When the dataset does not exist yet

Everything above is a dataset somebody already decided to build. That decision was made for a general market, which means the coverage boundary was drawn where the builder thought demand was, not where your question is. Sometimes the two do not meet, and no amount of shopping fixes it.

20. Forage AI

What it is. Forage AI builds the dataset when the shelf does not stock it. The engagement starts from the question rather than from a catalogue, identifies the sources that genuinely describe it, and produces a maintained feed in your schema. Forage AI delivers the data, not just the pipeline, and selector drift, anti-bot changes and schema changes are absorbed as part of the service instead of arriving as your tickets.

Best for. Three situations specifically. When the entities you care about are not securities and therefore appear in no market dataset: private operators, licensed entities, regional lenders, franchise networks. When the field you need exists only inside documents that nobody has structured. And when a signal you have been tracking manually needs to become a maintained series with a defined refresh. Every delivery passes a 3x QA team before it lands in your system, a 200% QA approach where every extraction goes through automated checks and human verification. On healthcare provider data, that process has produced a client-reported 99.7% field-level accuracy, which is the number we would ask any provider in this list to put in writing.

What it cannot do. It will not give you a price history this afternoon, and it should not. If a licensed dataset already covers your question, buy it, because a custom build is slower and harder to justify than a subscription that already works. The honest test is whether you have genuinely exhausted Bands A through D first. Our guide to financial data extraction covers what that build actually involves.

All 20 datasets compared side by side

Read the two defect columns first and the cost column last. A free dataset that carries survivorship bias is more expensive than a paid one that does not, because the cost arrives later and disguised as a strategy that stopped working.

Two names you will still see recommended that are not what you think

Both of these appear as live products on articles published in 2026, including the current top-ranking result for this search. We are naming them because following a recommendation to a product that no longer exists costs a day, and because it is a reasonable test of whether a list was researched or assembled.

Quandl

Quandl is not a product any more. It was acquired by Nasdaq and became Nasdaq Data Link, and in the process a large part of the free catalogue that made Quandl the default recommendation was withdrawn. The API model survived, which is why old code often still runs, and that partial continuity is exactly what keeps the name alive in articles and tutorials. If a guide tells you to get free financial data from Quandl, it was written from memory rather than from a browser. The current entry is Nasdaq Data Link, most of the catalogue is paid, and the licence attaches to each publisher rather than to the platform.

The Yahoo Finance API

There is no Yahoo Finance API and there has not been one since 2017. Yahoo shut the public API down that year, and everything since has been unofficial: libraries calling the same undocumented endpoints the website uses, with no key, no contract, and no commitment that any of it works next week. The tooling is genuinely good and the coverage is genuinely broad, which is why the practice persists. What does not exist is a licence permitting you to build a product on it, and that is a different problem from reliability, because no amount of retry logic fixes it.

How to choose: match the dataset to the claim you need to defend

Start from the sentence you will eventually have to say out loud, then work backwards to what the data must support. A dataset that is adequate for an exploratory chart is not adequate for a number in an investment committee memo, and the gap between those two standards is where most of the money in this category is spent.

# Companies that were real, and then were not.
DEAD = ["LEHMQ", "ENRNQ", "BBBYQ", "SIVBQ", "FTXMQ"]

for t in DEAD:
    rows = client.get_history(t, start="2005-01-01")
    print(t, "rows:", len(rows))   # zeros mean survivorship bias

Two checks are worth running before you commit to anything, and both take under an hour. The first is the delisting test: take five companies you know went bankrupt in your period, query them, and see what comes back. If they are absent, you have quantified your survivorship problem before it quantifies you. The second is the licence read: open the terms page rather than the pricing page, and find the sentence covering redistribution. If you cannot find that sentence, treat the answer as no.

The mistake worth naming explicitly is the free-then-upgrade plan. Teams prototype on a free source, get a result, and assume they can swap in a licensed dataset later. Sometimes that works. More often the free source carried a defect that produced the result in the first place, so the upgrade does not preserve the finding, it deletes it. Prototype on the free tier of the dataset you would actually buy, rather than on a different dataset entirely.

Frequently asked questions

What is a financial dataset?

It is any structured record of economic or market activity you can load and compute on, which in practice covers four different things that behave differently. Macroeconomic time series describe economies. Market data describes prices and the securities they attach to. Fundamentals and filings describe companies. Alternative data describes everything else that might carry signal, from news to search interest. The label is shared and the failure modes are not, so the first useful step with any dataset is deciding which of the four you are actually holding.

Where can I get free financial datasets?

The strongest free sources are official ones. FRED carries roughly 845,000 macroeconomic series from 126 sources and preserves the original vintage of revised numbers. SEC EDGAR holds every US filing as submitted, which makes it the only free source of as-filed fundamentals. The World Bank and IMF cover cross-country indicators, the ECB and Eurostat cover the euro area, and GDELT covers global news events at a scale no free commercial alternative matches. Free price data is the weaker category: Polygon.io, Tiingo and Financial Modeling Prep have workable free tiers, and the sources people reach for first are usually the ones to avoid.

What is the best dataset for stock market prediction?

The question is usually asked one level too late. No dataset predicts anything, and the ones that produce the most impressive backtests are generally the ones with the worst defects. For work you intend to trade, the requirements are specific: delisted securities present so the universe is not a survivor sample, corporate actions adjusted correctly so returns are real, and fundamentals joined on filing date rather than period end. CRSP with Compustat point-in-time meets all three if you have institutional access. EODHD combined with SEC EDGAR is the most practical route if you do not.

Is Yahoo Finance data free to use commercially?

No. Yahoo shut down its public finance API in 2017 and never replaced it, so the libraries in common use call undocumented endpoints with no agreement behind them. Their maintainers state plainly that they are unaffiliated with Yahoo and intended for research and educational use. That is a licensing problem rather than a technical one, which means reliability engineering does not solve it. Anything commercial needs a provider that actually sells you a licence.

What is survivorship bias in a financial dataset?

It is the distortion left behind when a dataset contains only securities that still exist. Build a universe from today's tickers, pull history for each, and every company that failed during the period has quietly dropped out. The scale is easy to underestimate: CRSP's survivor-bias-free fund database holds more than 64,000 open-end funds and over 31,000 of them are delisted. A backtest run on survivors overstates return and understates risk, and nothing in the file warns you.

Do I need point-in-time data for a backtest?

For anything involving fundamentals, yes. Two things go wrong without it. Joining fundamentals on the fiscal period end rather than the filing date gives the model numbers weeks before they were public. Using current values rather than as-filed ones gives it restatements that nobody had at the time. Both inflate results in ways that look like skill. Commercial point-in-time products solve this for a price, and SEC EDGAR solves it for free in exchange for the engineering work of rebuilding as-filed series from the filings themselves.

Related articles

• Financial Data Providers. The vendor-level view of this market, comparing who sells the feeds rather than what is inside them.
• Top Alternative Data Vendors. The full Band D roster, compared on dataset class, compliance posture and delivery.
• Top Investment Intelligence Providers. Twenty platforms grouped by the evidence layer each one proves, for buy-side workflows.
• Alternative Data Guide. The foundational explainer on source categories, use cases and how teams evaluate them.