When exploring the world of quantitative finance or algorithmic trading, you quickly end up facing a very common issue. Where do you get historical market data? If you have an account with Interactive Brokers, you can download historical data from them using Python. This article will show you how.
No matter what sort of analysis or trading you plan to do, you’ll need access to quality market data for your research and development. This can be a challenging and possibly expensive process. If all you want is daily U.S. Equity closing prices for large cap stocks, you’ll probably be able to find this from a number of free or close to free sources. However, you may want intraday data (prices at hourly, minute, or even sub minute levels). Or maybe you need data for other types of securities (futures, bonds, foreign stocks, for example). In this case, you will find the data to be a bit more expensive and difficult to find. For example, I found that historical 1 minute data for the full S&P 500 going back to 1998 will cost over $750 from several vendors.
However, some brokerages will give you access to historical data as part of their service offerings. For example, Interactive Brokers (IB) offers APIs for fetching historical data at different resolutions. For many, this data may be good enough for historical backtesting and research. And the price you are already paying for market data includes this data.
OK, there has to be a catch, right?
Yes, there are several issues with downloading data from a broker like IB. They point out that you may want to purchase your data from a vendor that specializes in historical market data. Some of these issues are:
- Being forced to use a clunky API instead of just downloading bulk CSV files. Some vendors, like Polygon offer bulk file downloads of historical data. IB forces you to use their API which adds a little bit of complexity to the process.
- IB has placed restrictions on their APIs to prevent users from abusing the system. Downloads should be rate limited to avoid being flagged for abuse of the system. IB’s servers will also rate limit your results if you send too many requests, and you may get disconnected.
- IB doesn’t offer historical data for stocks that are no longer listed. Your dataset will automatically suffer from survivorship bias. Some companies are acquired at high prices, others go bankrupt or are delisted. In both cases your historical backtests will have neither of these scenarios included. It also appears some expired futures data is not available, but I haven’t been able to verify this yet.
Even given these issues, using IB to obtain some historical data for research is worth considering as a first option. This is especially true if you’re already paying for the market data. If it doesn’t meet your needs, you can always purchase data form someone else.
In order to fetch historical data, you need to have met several criteria:
- Opened an IB account, and funded it
- Downloaded and configured the TWS software and python API
- Subscribed to Level 1 (top of book) market data for any contracts you wish to query
Along with these steps, IB places some limitations on fetching data:
- No more than 50 outstanding requests at a time. They note that it is probably more efficient to do fewer requests rather than try to test the upper limit.
- If asking for 30 second bars or lower, no 6 requests for the same contract in 2 seconds, 60 requests in 10 minutes, or two identical in 15 seconds. If you are grabbing consecutive single days for a symbol you can hit this limit pretty easily.
- In general, if your request will return more than a few thousand bars you should consider splitting it up.
So what sort of data is available? Bar data is available in sizes of 1, 5, 10, 15, and 30 seconds. Resolutions below 30 seconds are only available for six months from the current date. They will also generate larger bars of 1, 2, 3, 5, 10, 15, 20, and 30 minutes and 1, 2, 3, 4, and 8 hours, along with daily, weekly, and monthly bars. Those bars can consist of trades, bids and asks, midpoint, and various other fields described in the documentation. Note that building bars with last price and bid/ask will require at least two queries (TRADES and BID_ASK), then merging the data together. When considering the pacing of requests, this may factor into any downloading decisions.
In my testing, I found that more than a few thousand rows of data are returned for some queries (for example, fetching daily data for 40 years of AAPL returns over 9000 rows, 20 years of NVDA returns over 5000 rows at once. For minute bar data, I found that querying multiple days of daily data will cause rate limiting to take effect.
In order to run my code, you need to follow the directions from my earlier post to install the IB API. Once you’ve activated your Python virtualenv, you also need to make sure you’ve installed a few more Python libraries.
pyenv activate ib-example pip install python-dateutil matplotlib jupyter
I’ve posted a command line application to GitHub that allows for some flexible downloads of data. It supports a few different command line options for querying different ranges of data.
$ ./src/download_bars.py -h usage: Downloader for Interactive Brokers bar data. Using TWS API, will download historical instrument data and place csv files in a specified directory. Handles basic errors and reports issues with data that it finds. Examples: Get the continuous 1 minute bars for the E-mini future from GLOBEX ./download_bars.py --security-type CONTFUT --start-date 20191201 --end-date 20191228 --exchange GLOBEX ES Get 1 minute bars for US Equity AMGN for a few days ./download_bars.py --size "1 min" --start-date 20200202 --end-date 20200207 AMGN [-h] [-d] [--logfile LOGFILE] [-p PORT] [--size SIZE] [--duration DURATION] [-t DATA_TYPE] [--base-directory BASE_DIRECTORY] [--currency CURRENCY] [--exchange EXCHANGE] [--localsymbol LOCALSYMBOL] [--security-type SECURITY_TYPE] [--useRTH] [--start-date START_DATE] [--end-date END_DATE] [--max-days] symbol [symbol ...] positional arguments: symbol options: -h, --help show this help message and exit -d, --debug turn on debug logging --logfile LOGFILE log to file -p PORT, --port PORT local port for TWS connection --size SIZE bar size --duration DURATION bar duration -t DATA_TYPE, --data-type DATA_TYPE bar data type --base-directory BASE_DIRECTORY base directory to write bar files --currency CURRENCY currency for symbols --exchange EXCHANGE exchange for symbols --localsymbol LOCALSYMBOL local symbol (for futures) --security-type SECURITY_TYPE security type for symbols --useRTH use Regular Trading Hours --start-date START_DATE First day for bars --end-date END_DATE Last day for bars --max-days Set start date to earliest date
For example, to fetch all historical data for AAPL as daily bars and place the csv file in
./download_bars.py --max-days --size '1 day' AAPL
To fetch a week of 1 minute bars for AMGN, with each day saved as a separate csv file in
./download_bars.py --size "1 min" --start-date 20200202 --end-date 20200207 AMGN
You can refer to the code for more details. However, at a higher level using the IB historical data API involves several methods. First, I use the
reqHeadTimeStamp method to find the timestamp for the earliest data available for the contract. This is useful if we want to access the entire history of data, or to validate that we aren’t requesting data before the earliest date. Our result for this query is processed in the
headTimeStamp method. Next, we invoke the
reqHistoricalData method, making sure to request a reasonable amount of data. The results of this call are handled in the
historicalData method, which is called once for each bar. Once all the data has been delivered, the
historicalDataEnd method is invoked. There, we check that we’ve received all our data, save it to disk, and check to see if we have more data in our timespan to download. If so, we invoke the
reqHistoricalData method again, repeating this process until all the data is downloaded. All the IB methods are well documented in the IB API documentation.
I’ve also created a very simple Jupyter notebook that shows what some of the data looks like.