An update from me on my $50K weekly gap strategy. Apologies for not posting my weekend screen & uupdate yesterday - I built a new screener over the weekend and as a result there have been a few lessons learned, and also some consequences for this gap strategy:
Firstly, I completely re-built the screener in spreadsheet form (I had previously been using a script using Pro Real Time but couldn't overcome a particluar problem I was having with the screen). The screener is now totally spreadsheet based and is linked to my Norgate proce database which I subscribe to for my backtesting data. A couple of lessons I was reminded of:
I should use a screener that uses the same data source that I use for my backtesting
I should also have backtested my screener - apply it to a part of my backtest results to ensure it generates the same valid setup signals
I didn't do either originally, and so when I built the new screener and applied it to the backtest data I promptly found the answer to the question I have been asking myself since I started running the strategy live - the excessive number of version 2 gap setups being triggered (and generating losing trades).
What I found once I applied the new screener to the backtest data (a small sample of several months at least) - it was generating the right number of valid setups, but the number of setups triggered was higher than in my backtest data. When I examined why, I realised an error I had made in my setup definition - I will use the following segment of a version 2 gap setup to show what I mean:
The version 2 gap setup is defined as a sequence of more than 1 gap (at least 2, but there could be 3) occurring in the last 5 weekly bars. This happens relatively regularly (in both backtest and using the screener) but according to my backtest not triggering an entry very often (only 600 in 20 years, or about 30 per year), yet I had triggered 23 entries in the 1st month of live trading this strategy.
The gaps according to the definition could be any sequence of gap up opens or gap down opens. The above chart segment shows 1 such combination that illustrates why my live implementation had gone wrong. I had taken the entry in such setups to be where the entry for gap down 2 is, which is lower than that for gap down 1. But what I got wrong in my interpretation is that for the setup to be triggered in my backtest database, price had to break both gap entry levels (or actually just the highest, which according to the backtest spreadsheet formula is normally the earliest). This requirement obviously means fewer entries triggered since it requires a stronger price move to break both entry points shown above, and hence why a lot of the entries that were being triggered were losers because it required only a very weak price signal to trigger the lower entry level.
After reviewing my backtest database I also realised that while the entries in the database were triggered by the higher entry level, the price registered for the reult was the lower entry level, hence the version 2 gaps were showing a much better averge return than was actually the case. After correcting for the wrong price calculation, it turns out that the version 2 gap is actually not viable (which 1 month of trading so far supports), so I have decided to drop it from the strategy.
This means the strategy will now only comprise the version 5 gap (a gap down setup) and a version 6 gap (a gap up setup), which will reduce the number of trades that the strategy will take. I have closed the last of the version 2 gap trades on the basis they are invalid, and will continue to trade the v5 & v6 gaps, and post the progress here (I will update the trade results later to reflect this development).
A couple of additional lessons I've taken away from this:
I'm glad that I decided to trade the strategy at "half speed" (i.e. reduced weight, from a risk of $500/trade to $250/trade) for the 1st 6 months, as I anticipated initial teething problems executing the strategy (but perhaps starting at $100/trade might have been more prudent).
I need to modify my standard process to reflect the points above about developing, testing & implementing the screener needed for the strategy (this is the first time I've used a screener in my process, since I don't need screeners in my FX trading strategies, its all fully automated)
My instinct about there being something wrong was actually correct (it is satisfying knowing that I was right, even it ended up costing me money to find that out) - the warning signals I was getting from my comparisons of live performance to the backtest data proved to be right and worth paying attention to (even though I was slow in figuring out what the issue was, better slow than never)
The value (and effort) of measuring as much as I can in terms of strategy performance metrics, execution variance, etc, has definitely proven beneficial
It's a little embarrassing having posted this diary given the error I've made, but that's ok, I still feel happy to share my learnings, fwiw.
Hopefully from hereonin the v5 & v6 gaps will bring the strategy back into profit, although I know that could take some time, its about the process.
