Random Walk and Worst Case Backtesting, and Major Historical Events

Hi Cary,

Thank you for the input! We have data for US equities as far back as the beginning of 1998, and most of our other data stretches back to the early-to-late 2000s (options, forex, futures, CFD). We'll take this into consideration as something to implement in the future!

If you are interested in random algorithms, you should look here:

https://www.quantconnect.com/forum/discussion/1234/how-to-randomly-do-1000-trades-without-losing-money-risk-control-volatility-model-framework/p1

Partial solution using some hacks in the existing Lean repo:
As a QC user I took a step back from the web-based IDE and experimented with the Lean Github repo.  Found the Toolbox project withich is a CLI tool that can be used for many things but also for generating random walk data.  To make it work to create 29 years of random walk data for 15 fake symbols.  I ran into many issues and worked around them like this:

1. I made Toolbox project compile to x64 by changing its Build properies, so that all other assemblies that it loads which are compiled by default for All CPUs will be loaded as x64.  This was because I would run into OutOfMemory exception easily if I didn't.
2. I made a simple hack in RandomValueGenerator class so that it picked the same fake symbols every time:

private List<string> tickers = new List<string> { "SJX", "SKJ" , "SKX", "SNQ", etc 15 fake symbols};
private int tickerIndex = 0;

public virtual string NextUpperCaseString(int minLength, int maxLength)
{
//hack start
if (minLength != 3 || maxLength != 3) throw new Exception("Hack can't handle anything except 3");
if (tickerIndex > tickers.Count) throw new Exception("Hack not expecting more than tickers.Count");
return tickers[tickerIndex++];
//hack end

Then I could generate random walk data with a batch file full of lines like this:

echo y | QuantConnect.ToolBox.exe --app=randomdatagenerator --resolution=Minute --start=1991-01-02 --end=1994-03-21 --market=usa --data-density=Dense --include-coarse=true --symbol-count=15 --random-seed=395169 --ipo-percentage=0.0 --rename-percentage=0.0 --splits-percentage=0.0
echo y | QuantConnect.ToolBox.exe --app=randomdatagenerator --resolution=Minute --start=1994-03-22 --end=1997-10-27 --market=usa --data-density=Dense --include-coarse=true --symbol-count=15 --random-seed=395168 --ipo-percentage=0.0 --rename-percentage=0.0 --splits-percentage=0.0
echo y | QuantConnect.ToolBox.exe --app=randomdatagenerator --resolution=Minute --start=1997-10-28 --end=1998-08-13 --market=usa --data-density=Dense --include-coarse=true --symbol-count=15 --random-seed=395167 --ipo-percentage=0.0 --rename-percentage=0.0 --splits-percentage=0.0
echo y | QuantConnect.ToolBox.exe --app=randomdatagenerator --resolution=Minute --start=1998-08-14 --end=2001-09-17 --market=usa --data-density=Dense --include-coarse=true --symbol-count=15 --random-seed=395166 --ipo-percentage=0.0 --rename-percentage=0.0 --splits-percentage=0.0
echo y | QuantConnect.ToolBox.exe --app=randomdatagenerator --resolution=Minute --start=2001-09-18 --end=2008-10-20 --market=usa --data-density=Dense --include-coarse=true --symbol-count=15 --random-seed=395165 --ipo-percentage=0.0 --rename-percentage=0.0 --splits-percentage=0.0
echo y | QuantConnect.ToolBox.exe --app=randomdatagenerator --resolution=Minute --start=2008-10-21 --end=2011-08-15 --market=usa --data-density=Dense --include-coarse=true --symbol-count=15 --random-seed=395164 --ipo-percentage=0.0 --rename-percentage=0.0 --splits-percentage=0.0
echo y | QuantConnect.ToolBox.exe --app=randomdatagenerator --resolution=Minute --start=2011-08-16 --end=2015-08-31 --market=usa --data-density=Dense --include-coarse=true --symbol-count=15 --random-seed=395163 --ipo-percentage=0.0 --rename-percentage=0.0 --splits-percentage=0.0
echo y | QuantConnect.ToolBox.exe --app=randomdatagenerator --resolution=Minute --start=2015-09-01 --end=2020-03-20 --market=usa --data-density=Dense --include-coarse=true --symbol-count=15 --random-seed=395162 --ipo-percentage=0.0 --rename-percentage=0.0 --splits-percentage=0.0

I broke it up into time intervals between spikes of high volatility I saw in real world VIX charts because I had seen crashes and errors when trying to generate decades of random walk data.  I suppress the renaming, stock splits and IPOs in a hope of reducing risk of crashes and prevent the need for additional symbols.  To test my C# trading algorithm, I force Lean Launcher Debug build to run as x64, and run within the Docker container using the instructions here using VS 2019 (did not test with a Py algorithm, but I don't see why it wouldn't work unless x64 is an issue): 

https://github.com/QuantConnect/Lean/blob/master/.vs/readme.md

Conclusion:  I got my 29 years of random walk data with some huge jumps in the prices to simulate some sort of crisis.  Its only a step towards solving the issue.   

To address the original topic above, one would have to also hack the generation of price and volume data in such a way as to emulate more than a century of various depressions etc.

Hi Kevin,

Thanks for sharing your findings with the community. For anyone looking to generate random data, refer to this tutorial.

Best,
Derek Melchin