Knowledge in Market Data

Raymond Raud

July, 1997

Forecasting Topics Trading Markets Modeling with Neural Networks Timing Tools and Methods

Data is critical for any forecasting method or tool. Daily price, volume and open interest data is readily available from many sources. It includes market's internal trading patterns and reactions to the external events, but not the events themselves. Provided the external events occur at random they always remain an unknown factor.

The million dollar question in forecasting is how much predictive value is in historical price movement data? To answer the question the value must be defined in measurable terms. Those, in turn, depend on the trading system. To have something commonly understandable, I measure accuracy two ways:
- market move direction
- forecasted move's correlation with actual market move.
In these terms the best accuracy I have achieved is the direction of the market in 66 -- 76 % with forecasted movement correlation to the actual movement in 58% -- 65% range. Smaller forecasted moves are less accurate. The percentage is higher for trending futures (like indexes) and lower for sensitive agricultural issues.

From these results it appears clear, that the market is not random, it has its internal repeating pattern. Significant portion of the outcome depends on the external events that are random in nature. Forecasting from the price data will adjust to the reaction of the event from next period, but it is wrong for the period of the event itself. The following chart illustrates this conclusion on next day forecast for Live Cattle August 1997 contract. Two forecasting methods are charted together with the actual average price movement. One of the forecasting methods adjusts itself to the real price movement -- learns after forecasting. The other does not.

Both versions forecast basically the same pattern that generally follows closely the actual price movement. The difference is in some turning points (for example, 05/28/97 where the actual price movement turns one day before the forecasts) and also in additional change of direction that apparently is common for this market (for example, 06/23/97 and 04/30/97), but did not happen this time.

If my assumptions are correct, then forecasting accuracy can be improved by incorporating external event data. Selection of this is clearly individual for each market. For example, weather events are critical for agricultural markets whereas, interest rates probably influence index markets. Two problems are obvious from here:

What is more important: the markets expectation of the news (weather forecasts versus actual weather pattern, expectation of the Fed changing the interest rate or the actual act of changing it)
How to express this data numerically.

I welcome questions and discussion on any of my conclusions and assumptions.
Please drop me an e-mail, thank you.
Raymond Raud