This is for sharing ideas and for easy code sharing across forums.
Currently only holding the "FadeFinder" - a bare-bones backtesting script written in python.
It was in no way optimized for speed in version 1 (see notes below) but is now quite useful.
- 1) Improve speed - at first by implementing vectorization in the main function (version 2.0)
- 2) Implement commission and slippage - simulate realistic trading
- 3) Add a stop loss condition (static or dynamic) to limit downside, esp. important for small cap short positions
- 4) Add functionality to test long trades (currently only short positions)
- 5) Improve the measure of volatility - currently the script uses the average of entire testperiod thus "looking ahead"
- 6) Expand strategy evaluation - add e.g. Sharp/Sortino ratio and different levels of risk exposure
- 7) Add dynamic position sizing based on the historic volatility of each stock
The table of individual simulated trades 'resultsdataframe' and the eponymous 'Logdataframe' have the same structure as in version 1. The basic plotting and logging is also kept exactly as in version 1.
The easy solution here would have been to make it quite different, but setting it up in this way makes for easy comparison.
The difference lies in the main function which leads to...
Boosting performance (speed) was the main priority in version 2.0 so it is interesting to see the actual difference.
This test was done by using datetime.now() which is sometimes considered bad practice. The execution time is of such a magnitude that small inaccuracies do not affect the conclusions significantly.
Cons:
- It is not very exact
- The entire script is tested at once so it's not immediately clear what causes the difference
Pros:
- It is very simple
- It shows an actual usecase instead of a theoretic aspect and/or part of a larger whole
- Version 2 is much faster - due to vectorization of the main function
- There is a monotonic increase in execution time for both versions (not too much measurement variance)
- The time margin varies for both versions - the number of trading signals varies between different sections of the database
| Stocks tested | Version 1 | Version 2 |
|---|---|---|
| 100 | 20.46 | 3.90 |
| 200 | 49.59 | 6.28 |
| 300 | 75.47 | 11.45 |
| 400 | 82.77 | 15.62 |
| 500 | 102.14 | 17.43 |
| 600 | 131.92 | 19.30 |
When sharing version 1 I wrote: "Current performance on my Lenovo T450s: Going through 1800 stocks over 5 years of daily data takes ~8 minutes."
That was getting very bothersome and 2.0 runs through 4500 stocks in ~3 minutes. There is no reason to go back to version 1 other than for trying to understand the deeper logic behind Python. Performance wise it is too slow for larger datasets and version 2 leads to the exact same data.