Forex Automaton

It's now official: from now on, a forex forecast of low, high and close for the next 24-hour period will be posted on this site daily at 9am Eastern time. The system is named Danica following the naming convention where first names starting with D are assigned to systems with daily decision-making scale.

Danica's forecasts are of much wanted one-handed nature: there is no "on the other hand" caveat. Mathematically, this is a projection of what might happen on the subspace of what used to happen under similar pre-history conditions, if such an explanation is of any help to the user. This is about as much as I can say in this black-box, closed source project. The result of such a procedure is by necessity a number -- a forecast of the next day's close, low and high for each of the 14 exchange rates followed.

This announcement will be followed by one or several more detailed posts dealing with the optimization trade-in chosen for Danica. These will be placed in the Forex trading system: are we there yet? section of the site.

In brief, trade-ins had to be made to resolve the following dilemmas:

• What is the best for high and low is not the best for close. The choice made is in favor of predictability for close and to the detriment of predictability for high and low. Hypothetically, one could imagine a trader reading our daily forecast and trading intra-day relying on our predictions of change for high or low. Those predictions can have historically much better correlation with reality (35-40%, see Prediction quality for high and low...). Naturally such a user would want to have the highest possible predictability for high and close. However, it appears that such a performance can only be delivered at a price, and the price is reduced usefulness of the forecast for close. Namely, the optimized parameter value best to predictability for daily high and low was seen to land in the area where the system was seen to be of little use for close except under conditions of extreme volatility. This would imply increased risk for our hypothetical high/low user. Therefore, a compromise solution is chosen where the predictability for high and low is reduced a bit (from 35-40% to around 30%) but in return, the predictability for close is expected to be robust -- which is in the interests of the users who will choose to work with the system strictly on its nominal decision-making scale: day scale, or close to close.

• What is best for one forex pair is not (within the available time period of observation) necessarily the best for every other pair. One path could be to optimize for every forex pair separately. Such a hypothetical solution was most decisively rejected -- it is suspected to lead to Mexican-peso-type scenarios (in Nassim Taleb's usage of the term). Instead I decided to err on the side of less fine-tuning in hope that the resulting solution will be of higher value in the long term. It's worth emphasizing that the system is optimized for the 14 exchange rates simultaneously with a single adjustable parameter -- and while success is limited (4-5% predictability correlation for close), there is also a chance that this way, something truly archetypal to the forex markets in general is captured by the forecasting model, and whatever single solution works on average across the exchange rates without ad hoc adjustments, might as well be expected to work throughout history.

• What is best for high volatility regimes is not necessarily the best for any volatility. An attempt was made to avoid a system useful exclusively in the extreme volatility regimes; even though the largest contribution to the figure of merit chosen will come from the volatile episodes, the system is expected to at least remain meaningful in the every-day environment, and to be of net benefit.

A feature of Danica approach is that along with the forecast, moving averages of a certain figure of merit, characterizing the quality of prior forecasts, are published. This figure of merit is the quantity referred to in the paragraphs above as "correlation with reality" and "predictability correlation". In strict terms, this is a Pearson correlation coefficient between the actual and predicted logarithmic returns. There is no bias coming from the selection of days entering the summation: for a specified range of a moving average -- week (6 trading days), month (24 days) or since inception -- all days falling within the range are used.

The system does not run a model portfolio and does not tell you whether a given situation is a "buy" or "sell", thus a concept of risk can not be associated with such a system -- only the measure of information quality such as the one just explained. Those who naively believe every "up" forecast is a buy advice and every "down" one is a sell are very far --dangerously far -- from the truth: trading history simulations with algorithms just like Danica indicate that one can fail even with very good forecasts if other components of a trading system -- such a stop loss and a trade idea discrimination -- are not properly thought through. And as if the legal disclaimers were not enough, let me underscore that all responsibility for the ways you use this information -- which is currently made available publicly free of charge for your entertainment only -- rests with yourself.

Bookmark with: