## Introducing Heidi, the hour-scale predictive model

Written by Forex Automaton
Wednesday, 11 August 2010 11:39

ForexAutomaton has just expanded its portfolio of free predictive models into higher frequency domain. From now on, a forex forecast of low, high and close for the next hour will be posted on this site every 60 minutes. The new system is named Heidi following the naming convention where first names starting with H are assigned to systems with hourly decision-making scale.

#### How it works

As has been the case with Danica, Heidi's forecasts are of the much wanted one-handed nature: there is no "on the other hand" caveat. Mathematically, Heidi's forecast is a multidimensional 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 one 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.

#### Heidi's initial learning is barely over

Compared to Danica, Heidi operates on the hour scale with 24 times more statistics available to draw conclusions of very high statistical significance. Therefore we know very well what to expect. To keep it fun, Heidi started accumulating its figures of merit on 2010-07-29 18:00 (all dates are in Eastern time unless specified otherwise). This is the date of inception referred to in the "since inception" column of the system output. The relative weight of the back-testing history, even in the "since inception" figures of merit, will diminish as the time goes on, and will become insignificant fairly soon -- which will hopefully make the experience of watching the evolution of these figures of merit a lot more fun.

#### Figures of merit: what they are

Our chosen figure of merit is Pearson correlation coefficient between the predicted and real logarithmic returns.

By construction, the Pearson correlation coefficient is a quantity bounded between -1 (the forecast move and reality are total opposites) and 1 (the forecast is perfect). A success or lack thereof for every hour makes a contribution to this quantity.

The coefficient is based on the product of the predicted and actual log returns. The product provides the desired incorporation of the signal strength, not just direction, into the figure of merit. A hypothetic rational operator of the system will not pursue small forecast moves, understanding this to be a noisy system. In order to make a large positive contribution, one needs a coincidence of a large move in a currency pair with a large forecast move in the same direction. This statement can be mirrored for a large negative contribution, while the rest of cases fall in between these two extremes.

#### Figures of merit: what to expect

Based on the high statistical power of initial tests, the correlation coefficients for high and low are expected to converge to values in the vicinity of 0.3 (30% positive correlation). The long-standing challenge of this time scale has been that the correlation coefficients for close are expected to converge to values in the vicinity of -0.03 (3% negative correlation). One artificial way of obtaining positive correlations for close could be to take the code apart and flip the sign of the predicted return for close. Had we done it, a typical system output would predict contradicting directions of change for low and high on the one hand, and close on the other. On the surface, that would have solved the problem but we decided to leave everything as it is.

In the course of the optimization, the adjustable parameter responsible for forecasting was placed at the value which maximized the positive correlation of the forecast with reality for the hourly high and low, while ignoring the effect on close.

#### Figures of merit: what does it mean?

Positive correlations of forecasts with reality of the hourly extremes imply that a user gets an edge (a bias towards success) by placing the stop loss on the previous hour's extreme, expected not to be reached (low  for long trades, high for short), and trading in the direction of the other extreme which is expected to be exceeded (high for long trades, low for short). The "negative expectancy" forecast for close will serve as a reminder to the user to try not to wait for the close to arrive.

This is consistent with the mental picture of a market which is self-correcting in an excessive way.

### Bookmark with:

Last Updated ( Tuesday, 29 January 2013 15:56 )

### World Time

New York: 0:34 London: 5:34 Berlin: 6:34 Tokyo: 13:34