Forex Tri-Arb signals as predictors of price

[Talha123]

Yesterday I was checking out this thread at FF –Triangular Arbitrage @ Forex Factory– and I found it extremely interesting.

So the story is this: for any forex currency cross (say, EURJPY) there are combinations of two other currencies (for example, USDJPY and EURUSD or EURGBP and GBPJPY) from which you can create a “synthetic” EURJPY. So if the synthetic EURJPY created from USDJPY and EURUSD deviated substantially from the actual EURJPY, you would have a risk-free arbitrage opportunity because of a thing called the Fractional Product Inefficiency. Of course, that never happens for us, because of spreads+slippage, the institutional guys get it first at microseconds speeds, etc.

But what if we could use those deviations as a way to predict where the actual price of the base currency -EURJPY in this case- was going? That’s what the guy at FF was doing very successfully, and he doesn’t seem like a scam or something. It seems pretty legit. He takes into account the different markets open and close times so he can distinguish the actual currency movements from the movements of the elements that constitute the synthetic currency.

So I threw together an EA as a proff of concept and it makes money in the backtester bot only for very small spreads -I tested with 0,1 pips-, for 2 or 3 pips it generates losses with the parameters I tested. So is suspect those were non-exploitable inefficiencies. Anyways, the EA that guy at FF was using as he described it looks more complicated than what I described here, this is just the basic premise. The catch is that he doesn’t publish code or detailed explanations, just some There’s also a nice indicator floating around from another forum (mentioned in the FF thread) that shows the synthetic prices created from various rings, but it doesn’t plots them and it doesn’t works in the backtester.

I don’t trust the MT4 backtester anyway (it looks like he tested his system live and it worked live, so the results in the FF thread are not necesarilly BS), and this is supposed to be sub M1 so now I’m playing with Dukascopy’s data in R to see if I can get some insight into this hypothetical inefficiency. But I have exams coming up in the following weeks so I can’t dedicate a lot of time for this, if there’s any interest to contribute to the idea, I’ll share what I find and try to uplaod the little code I wrote.

But the main thing I’m wondering is… Which are the probabilities of this being exploitable, and which of me just wasting my time? Opinions?