Maybe this is useful, so I wanted to share it with WL community. I am using the Analysis Series tab under Backtesting results, in order to learning about different indicator capabilities and what to expect. I backtest a Buy at Market, Sell at Market 3 bars after, at different scales using a popular ETF in order to compare indicators predictability. Very interesting results. See the following image:
It shows us RSI can predict profit in resolutions of one week to a month, at a day scale it barely filters some false alarms, lesser than this it looses quantitative resolution due to low signal/noise.
If you like this post, comment and contribute with ideas and your own experience, it will help us to learn, to improve our tradings and create even better strategies!
It shows us RSI can predict profit in resolutions of one week to a month, at a day scale it barely filters some false alarms, lesser than this it looses quantitative resolution due to low signal/noise.
If you like this post, comment and contribute with ideas and your own experience, it will help us to learn, to improve our tradings and create even better strategies!
Rename
This is the kind of analysis that I love to see. It would make a great basis for an article at a publication like Stocks and Commodities magazine. You should consider fleshing this out and submitting it.
Some things to flesh out for your TASC article :)...
1. << I backtest a Buy at Market, Sell at Market 3 bars after >>
If it wasn't done, the test should be set up for a Multi-Position backtest to make sure you're testing each bar.
2. Is there a relationship between the RSI period and the number of days of expected profit. You only tested 3 bars with a 14-bar RSI.
3. What is RSI doing for each data point? Turning up/down? Consec bars down/up? See the Indicator Profiler for more clues.
4. What is the trend of the underlying and how would you define the trend? There was a TASC article years ago that pointed out that when an instrument is in a strong rising trend, the 14-day RSI rarely dips below 40, or some rather high value.
You concluded that it "barely filters", yet I can certainly show you a study that RSI is a good filter for values below, say, 25, depending on other factors. (see above)
1. << I backtest a Buy at Market, Sell at Market 3 bars after >>
If it wasn't done, the test should be set up for a Multi-Position backtest to make sure you're testing each bar.
2. Is there a relationship between the RSI period and the number of days of expected profit. You only tested 3 bars with a 14-bar RSI.
3. What is RSI doing for each data point? Turning up/down? Consec bars down/up? See the Indicator Profiler for more clues.
4. What is the trend of the underlying and how would you define the trend? There was a TASC article years ago that pointed out that when an instrument is in a strong rising trend, the 14-day RSI rarely dips below 40, or some rather high value.
You concluded that it "barely filters", yet I can certainly show you a study that RSI is a good filter for values below, say, 25, depending on other factors. (see above)
I really don't want to remove the fun from the game, but after all we want results which lead to a robust, profitable strategy. Therefore....
Caution:
Please make sure your data points are statistically significant, i.e. there are enough trades behind a single (result-)number.
After years of research I came to the conclusion that a single data point (average, average profit, etc) should be calculated from at least 100 "trades" to make it useful.
Else all you see are "statistical fluctuations" without any merit for realtime trading.
The other big question: What happens "Out of sample"?
In an analysis like the one shown above this means: You should observe/show results for at least two independent time intervals like 5 years (named IS) and the following five years (named OS) to make sure the effects are stable/robust over time.
I used to work with four(!) time intervals which enabled me to reject any effects which are not present in at least three out of four intervals.
Caution:
Please make sure your data points are statistically significant, i.e. there are enough trades behind a single (result-)number.
After years of research I came to the conclusion that a single data point (average, average profit, etc) should be calculated from at least 100 "trades" to make it useful.
Else all you see are "statistical fluctuations" without any merit for realtime trading.
The other big question: What happens "Out of sample"?
In an analysis like the one shown above this means: You should observe/show results for at least two independent time intervals like 5 years (named IS) and the following five years (named OS) to make sure the effects are stable/robust over time.
I used to work with four(!) time intervals which enabled me to reject any effects which are not present in at least three out of four intervals.
Your Response
Post
Edit Post
Login is required