Okay so the last results were promising but I think I’m getting ahead of myself.
I want to find Indicator 1 (I1) that predicts if the market conditions are good for my system for above equilibrium Strike Rate and Indicator 2 that tracks the performance of I1 and lets me know effectively if I should trust its advice. I started I1 & I2 using 2 Stochastic Oscillators and randomly changing the Trigger criteria to see what worked best.
What I really ought to do is record the strike rates and feed that data into the formula for all the trend indicators and oscillators and systematically identify the correlation between the probability of wining the current trade and each variable in the respective indicator or oscillator (for Indicator 1) and similar for I2.
Once I have the correlations I can select variables with the highest impact and feed them into a multiple linear regression model to find how they interact to determine the likely strike rate/Win Rate for the current trade (for Indicator 1).
So for I1 for example I might look at the correlation between the absolute value of the last 20 strike rate (SR) to the probability of winning on the current trade, or the absolute %K reading on a Stochastic Oscillator, %D “…”,%K(1) - %K(0),%K - %D, SR - MAofSR etc…
So I will take 18 currency pairs and break them down as follows:
6 Pairs: Training Set: The data from these pairs will form the basis of the initial correlation model
6 Pairs: Validation Set: I will test and improve the correlation model using this data
6 Pairs: Test Set: Will test the modified model on completely new data to see if the results stick.
The Training Set is:
- AUDUSD
- CADCHF
- EURAUD
- EURCHF
- EURGBP
- EURJPY
Altogether the win lose sequencing gave me circa 51,000 trades to work with, from 4th Jan - 2nd Sept 2016.
Starting the correlation model with the simplest variable - the strike rate (SR) in the last 20.
There were multiple readings for 0% SR in the last 20 but the highest was only 75%, funny how that works in what should be a completely random system; I think it might have something to do with the spread which changes the R:R balance in the distance traveled to win versus the distance traveled to lose.
Y-AXIS (%Wins) || X-AXIS (Last 20 SR): I was expecting a horizontal (0 correlation) line here but the results were surprising; positive correlation, going from 55% SR to 75% SR the probability of wining the current trade was consistently above 30% and sloping up.
Okay so, I should really fit the model properly and then move on to the next variable, but I couldn’t help but write a quick EA to test it.
The EA enters 0.01 Lot trades at random with an R:R 1:2.5 with the recorded outcomes Win, Lose (1,-1). After 20 records start calculating the SR for the last 20 trades, if the 0.55 <= SR < 0.8 then on the current trade increase the Lot Size from 0.01 to 10 Lots (risking 0.71% of account balance per Trade).
Note: I should spread the risk to multiple currency pairs to increase the trade count so that the results tend to the true mean.
Results from 4th Jan - 2nd Sept 2016
USDCHF || 4,495 Trades || -£12,977
GBPUSD || 11,981 Trades || +£41,168
EURUSD || 15,756 Trades || +£41,168 (The same: odd)
USDJPY || 9,743 Trades || +£30,026
USDCAD || 10,147 Trades || +£3,582
AUDUSD || 4,723 Trades || +£10,376
EURGBP || 5,419 Trades || -£14,202
EURAUD || 23,041 Trades (High?) || -£8,332
EURCHF || 4,354 Trades || -£4,349
EURJPY || 12,171 Trades || £32,941
Total = £119,384. Initial account balance was £100,000 (roughly 0.71% risk per trade).
The last 20 SR seems to be a good indicator for future performance to produce a win rate above the equilibrium value for +EV, there are still many other variables to process and I am quite surprised that this appears to work on something as straightforward as that.
(I suppose this could be gamblers fallacy operating in reverse i.e. for a high consecutive win count - and I suppose it does make make sense in common sense terms of getting in when the recent performance is higher)
Should have really recorded the lowest account balance to get a better idea of the downside risk and other stuff, but it was just a quick test to see if the results make any sense in reality; though the absolute draw-down was consistently low and especially tight for the pairs that ultimately lost.
If anybody is interested please help me test the results too just to make sure I haven’t made a computing error somewhere.
Simple 20 SR Random.zip (11.6 KB)