We started by building a support vector machine model based on features used in harmonic trading, with the idea that ideal “harmonic” ratios can be learned rather than explicitly specified. This worked on testing sets but not when we started trading with it. We abandoned the model before we realized that we need to manage risk and expectations through the use of stop loss and take profit orders, and before we realized that large financial institutions sometimes instantaneously “game” the price flux to knock out smaller players’ positions.

Fortunately all this occurred through a practice account, so we paid no price for our arrogant entry into the currency trading arena other than facing a bit of humble pie. We realized we needed to learn more, and therefore read the books [1], [2], and [3] to understand as much as we can about FOREX and technical analysis. We still constantly revisit these sources to improve our insight.

We remained at this point committed to a technical analysis-based approach, one suitable for automation, rather than to adopting a manually intensive fundamentals-based approach.

Our next strategy applied Autochartist’s predictions available from Oanda, our broker. This improved our results, as in our losses lessened, but the Autochartist data does not forecast ideal stop loss and take profit positions. So we experimented with everything between simple linear regression and ANOVA to come up with a stop loss boundary—one that we updated programmatically over time. ANOVA worked best but proved too computationally demanding for the EC2 instance we selected (for budget reasons) for this project.

Then we tried identifying through an SVM model turnaround points in the price time series, with the idea that if we correctly identify them most of the time while setting appropriate stop loss and take profit boundaries, we will succeed. The problem was that the SVM features we used only looked about five periods back (or were computed from these five periods). So unlike the ANOVA, it did not take into account the greater dynamics of the time series.

We are about to try deep-learning the time series dynamics using about 500 periods back of information taken together, and using it to forecast the next 18 or so periods.

# References

- Currency Trading for Dummies. Kathleen Brooks and Brian Dolan. 2015
- Technical Analysis for Dummies, Barbara Rockefeller. 2014
- Forex Survival Manual, Save Your Trading Account From Collapsing. Salman Shariff. 2016. e-book.