The New Era in Property Price Forecasting

Factors affecting future property values are partially reflected both in their past prices and those of related financial instruments. Influences of those factors can be distilled through careful mathematical analysis and built into a non-linear forecasting model. TFL has been closely associated with the development of this method, has implemented it for a growing number of well-recognized clients across the U.K. and now seeks to make quality forecasts more widely available throughout the property profession.

TFL pushes the forecasting horizon well beyond three years on a monthly basis. This is a significant improvement over other methods of technical forecasting which previously had a horizon of only a few months or were based on linear modelling techniques.

The Mathematical Specifics

A more detailed look at TFL's forecasting process can be best understood if discussed in reverse order of their use in the process. In this spirit, we shall begin at the end.

Final Forecast Product

The final forecast is derived from a statistical average of a larger number of statistically independent models. This is a result of employing Radial Basis Functions (RBFs) and proprietary model averaging techniques. RBFs are a specialised computational method in Neural Computing.

The true power in employing radial basis functions is two-fold.

1. Significantly improved computational speed
2. Highly stable forecasts over longer forecasting windows

Independent Models

Each independent model is created from what appears to be a simple combination of the target series and an associated, parallel series that has been determined to share mutual information with the target series. In this way, the number of simple models created are limited only by the number of parallel series and the number of initiation points for each target/parallel series combination.

This is an important development since TFL follows the auspices of Occam's Razor which states that the best model is as simple as possible.

TFL forms very simple forecasting models using Radial Basis Functions to create step-wise forecasts . Then through the prudent and knowledgeable application of proprietary averaging techniques, uses these simple models to create longer term, and more accurate forecasts than could otherwise be possible alone

Finding the Relevant Parallel Series

It has long been recognized in any form of data analysis, the selection of appropriate inputs is crucial in obtaining a relevant output for the final forecast. Subsequently, TFL has developed a proprietary application from the field of Information Theory that identifies the most appropriate series to include in any particular forecast.

The question of what defines 'most appropriate' is the critical key to the selection of the parallel series. TFL's methods ensure that the parallel series to be used in the forecasting models are as statistically independent as possible as well as sharing a high degree of mutual information with the target series.

TFL uses 100+ parallel series that are tested against EVERY model to determine their contribution in the sharing of mutual information for the target series. Over 1.5 million forecasts are performed on a monthly basis and distributed to TFL clients.