Live Profitable FX Quant Algo-Trading System

*backtest data is at the end*

Methodology:

I must first say a little about myself and my approach to trading FX, first things first, I am a an experienced programmer (I began at age 11 and am now 16),
I write my code primarily in Java, and have developed a backtesting infrastructure with data from Quandl as well as a parser to read my brokerage's historical
FX data. I follow the assumption (based on evidence) that a single asset's past prices alone cannot predict that of the future, however when external information is applied (econometrics, other asset prices and correlation, order flow/volume, as well as implied volatility) all can be used to forecast future performance. An example of a system I may develop (for explanation purposes) would be short the dollar if Non Farm is below 150k, other wise hold. Compared this to others who may use
technical indicators which I have no data suggesting they are capable of predicting the market.

Current situation:

At the moment I have two strategies trading, these are swing trading systems making trading decisions based on fundamentals, each having four hours of exposure unless I intervene due to reasons such as FOMC press conferences/speeches or other news events that causes the trade to potentially reverse. The systems do not have stops/takes because I have found that greater returns are possible without them and as well due to how I mathematically model the effect of these events on volatility. My current goals are to build a portfolio of 5 market neutral strategies all with a low correlation to one another's returns, 3-4 short term, as well as 1-2 long term. I also plan to target a sharpe ratio of 2.5 or higher. Sharpe ratio for those who don't know is mean returns divided by the standard deviation of returns. A rule I have set for my short term strategies is a max leverage of 1:10, and 1:4 for long term. My strategies are currently very low latency, with only two opportunities a week and each trading an average of 40% of weeks (this means I will make a trade 64% of weeks). Finally I expect to make at least 5% return a month and will be compounding monthly hoping for 110%-200% return a year..

Another thing to note about backtests is that there is only 3 months of 30 minute bar data on my hard drive, If anyone can get me more data that is from Oanda that would be phenomenal!!

(Backtest Data Assuming Leverage)

Strategy A Statistics:
CAGR: 52%
Sharpe: 2.85

Strategy B Statistics:
CAGR: 60%
Sharpe: 3.01

Strategy A Graph:


Strategy B Graph:

UPDATE:
A trade was made today, the first since starting the models live, it was strategy A, I intervened with a USD/CAD position which lost 0.4% however had I not intervened I would have made money on it. I also made 3% on a USD/JPY trade, and if i didn't intervene i would have made even more.

moral is that I should have more faith in my mathematical models.

I suggest you scale down to 0.01 while you're testing so you aren't tempted to close your positions from drawdown. This way you'll have a clearer picture of the veracity of your system.

Yes I would agree except have paper traded the model prior twice (just looking at the morning price, waiting later in the day to verify). It's just that waiting till the end is what is difficult to do and now that I have a feel for it live I believe it will become easier.

UPDATE:
I have sampled strategy A's probability of a favorable bar (ie if long up bar is favorable), and have also sampled the mean movement of both unfavorable and favorable bars separately. Because I only have a lengthy amount of 30 minute data and wanted minute by minute data I have statistically adjusted to minute by minute data using weightings as shown:

note t is time >= 0
p is probability vector containing the chance as a percent that a bar moves favorably

This makes the time series continuous and different for each minute bar simulated. I then model the paths using a monte
carlo method which simulates numerous outcomes that could possibly occur (in this case 250 possible outcomes) which gives me a better idea of the likely hood of generating a certain return.

The reason this is helps me almost more than my back test did is because my backtest only consisted of 5 actual trades, which brings up another concern. Is 5 trades a sufficient sample size? No is the answer however I had backtested this model over a daily interval and had managed to see 15% unleveraged returns when holding trades for a week at a time (with significantly lower sharpe ratio)

None the less, here is the simulation results for 250 different outcomes (pls note that the time the trade is open is 4 hours, the return is in percent, and the bars are 1 minute):

Hello subscribers and new viewers, I attached a poll so please vote, I am just curious c:

UPDATE:
Today Strategy B made a trade short the dollar, which I took out early due to trade balance out performing expectations and an FOMC speech, and though that limited the profits, it did reduce risk and possibly lock in gains. A return of 1.1% was still achieved which sets my CAGR at 101.4%. Basically what this means is that my return should be 101.4% in a years time if I do not compound and continue at current pace. This is very impressive given the relatively low risk of my models. I must also state that Strategy A could possibly fill a trade tomorrow which could potentially boost my CAGR even higher if it is a successful trade.

I also recommend filling out the poll as I am eager to hear your opinions!

UPDATE:

Hello and once again i'm back! I have been working hard at creating a new back testing framework and in the process of doing so I have developed a new long term trading strategy. This strategy implements machine learning for stat-arb and could see returns as high as +100% after leverage but is yet to be tested live as I haven't finished programming it ;/ However I still want to show my progress! It took a while to backtest the machine learning versions (28 fricking minutes!) However it was worth it!

Pair 1 backtest no machine learning:
CAGR: 8%



Pair 1 backtest with machine learning:
CAGR: 12%



Pair 2 backtest with machine learning:
CAGR: 31%

Hi!

Not sure if I understand your system but why no trade today for example with the bad NFP news?

All the best, I think you have a bright future infront of you

Ahh yess, I have prevented my systems from trading on Thursday and Wednesday of this week because of FOMC speeches occurring around the time that the trade would be made. As for NFP, my strategies do not trade Payrolls. Thank you for encouragement! I also plan to launch an intraday stat-arb strategy on Sunday night.

I'm now going live with a new strategy for my portfolio. This strategy makes intensive usage of machine learning in the form of a genetics optimization in order to adapt to it's environment. This particular strategy uses correlation and co integration to model how various currency pairs move together in order to spot inefficiencies that can be exploited to make a profit. This strategy does not keep an order open based on time of day or specific price points but rather it keeps an order open based on divergence in correlation and co integration between two pairs. This means that losses are possible, however the algo is still correct 100% of the time as it does not care about price but rather the movement of pairs relative to each other. The strategy can have a trade open for any amount of time, however 6 - 36 hrs is an expected trade time, thus my algo will run with low leverage to prevent losses. I can still have fairly high leverage however because of the fact that my strategy is hedged, so despite starting it with 1:1 leverage, I intend to scale up to 4:1. Below is a back test over a three month time frame.

CAGR: ~40% with no leverage ~160% with leverage



If you have Questions feel free to ask me

UPDATE:
I have added a TakeProfit to my newest algo which is pre calculated in hopes of boosting performance.

UPDATE:
It appears my system would reach a high point typically of 12-15 pips, before drawdown would bring the pl negative, I have cut back my calculated TP using a scaler quantity of 0.25 should boost success rate, but will limit gains.

Update: I have decided to stick with 4 daily timeframe stat arb strategies for the summer, they make heavy usage of machine learning through a genetics optimization. I would use my initial two strategies except for the fact they must be monitored and my sleep schedule is outta wack (it's summer man lol).

My new thread: http://www.forexfactory.com/showthre...66#post9148966