I'm about to embark on training a neural network on daily forex data, with a view to obtaining a predictive network. I'm also interested in using data other than the forex currency pair data itself, in a manner similar to intermarket analysis. What other time series data does the panel think will provide meaningful input? Obviously, various other forex cross rates are important, along with perhaps interest rate time series. But what about perhaps less intuitively obvious time series? I'm more interested in time series that have a justifiably fundamental reason for inclusion rather than those that might simply exhibit historical correlation. Links to online references/papers, e.g. SSRN etc. would be very welcome.
In addition to FX currency cross rates and interest rates, several other potentially useful inputs are:
1) Economic data (GDP, inflation rates, employment figures) for the specific countries whose currencies you are interested in. While these data may be useful indicators, there are however two problems: Firstly the granularity of the data. If you are using EOD or intra-day FX data then ideally you would like the other inputs to your NN to be on a similar timescale and these are not. The second problem is that government statistics are often subject to "adjustment" some time after issue. Notwithstanding these caveats, such economic data may be useful with regard to long-term trends.
2) FX rates are relative values of one fiat currency against another without any absolute scale of "true" value. It is useful to also include as input "hard" asset price series such as precious and base metals and energy. These help to provide at least some form of absolute reference. For example a rising price of gold (denominated in USD) can alternatively be viewed as a decline in the value of the USD vs. hard assets.
It's also necessary to look into technical indicators and filters. Technically analysis are often widely employed by finance practitioners and can apply to any sort of timeframe whether that's intra-day or EODs. Since ANNs are able to fit complex non-linear inputs, it would not hurt to add many of those indicators into the mix
Actually you can find a papers talking about some relationship between almost any two types of indicators. But based on my work this what I suggest you add:
- Commodities futures (continues) (Many paper on the relationship between oil prices and USD, gold/silver and USD)
- Market indices (DJI, S&P500, DAX, SET, NZ40, ...)
I would also suggest that you inverse the patterns of these indices and futures and add both original and inversed pattern.
Interest rate differentials is the most justifiable fundamental input, also the most explanatory second order input you should use. Another is the slope of the yield curve and the difference in slope between pairs. You can also look at the relative difference in speed of interest rate changes. Another fundamental input is the real IR spread. Good luck.