Frequently Asked - Technical Questions

What are your methods?

We have a structured framework, optimised for individual target time series, to provide a rational procedure for the forecasting process. Preliminary mathematical transformations are applied to maximise the strengths of relationships between current and past values of the target time series. We examine other time series to see which are most closely associated with the target series, and include the most suitable ones in our models. Forecasting is achieved with complexity-optimised non-linear models. Finally the original mathematical transformations are inverted to produce the necessary forecasts.

The unique feature of the service offered by TFL is that our forecasts are the most likely scientific projections of past patterns and current influences into the future, using only existing data. Our forecasting processes use the powerful methodologies of Neural Networks in the form of Radial Basis Functions and Bayesian statistics, a combination which indicates the most probable future direction of data, and which can thus forecast for up to five years ahead.

How much of a forecast is based on economic opinion?

None. Zero. The TFL process is completely free from opinion or subjectivity either economic or otherwise. This may be the single biggest advantage to the TFL offering since this method provides a bias-free perspective.

Please note that this does not mean that TFL forecasts ignore economical data! Parallel series with economic origin are always used as part of the TFL process. The main differentiation is that it is free from opinion.

How do you measure performance?

TFL does not measure performance other than to periodically verify the accuracy of our forecasts by running historical tests. This serves to add a dimension of validation to the methods we use.

If you have no economic expert, how do your forecasts compare with other industry experts?

TFL clients have evaluated some of our forecasts against their own methods. In a great many cases, TFL forecasts have confirmed their economic-based forecasts. Most importantly, TFL is able to recognize significant features such as turning points in specific markets further into the future than is possible with current econometric methods.

What happens to your forecasts in instances such as earthquakes or major disasters?

This is known as a shock to the data series. It often creates a discontinuity in the data series that causes most forecasting models to become completely ineffective. TFL's methods recover more accurately and more quickly than any other methods known to us.

Why are Proportionately Complex models so important?

• If the model is too simple, the true nature of the data is not properly represented.
• If the model is over complex, the model tends to represent the noise rather than the true sign within the data.

Read more about Occam's Razor

Why are Radial Basis Functions so important in forecasting?

RBFs, when employed with the proper SKILL, will give very robust and stable forecasts further into the future than any other method. RBFs alone are not the sole basis for the success of the technology employed by TFL.

What other factors are important to TFL's success?

1. data preparation
2. selection of the appropriate parallel series
3. prudent use of neural network technology
4. signal to noise decomposition
5. proprietary model averaging
6. error minimisation
7. efficient computation methods
8. proper and efficient use of computer hardware

Your forecasts are only as good as the models upon which they are based?

True. And, because TFL's methods are soundly statistically based, TFL can reasonably claim that due to model averaging, the cumulative error in our final forecast is extremely small and can be expected to become smaller as more models are recognized as associated to the target series.

Your forecasts can only be as good as the forecasts of the parallel series you use to forecast the target series.

Yes, and because the parallel series are forecast at the same time as the target series, in a stepwise fashion, the combination of statistical methods with some advanced proprietary mathematical techniques allows each forecasted model to be as accurate as possible based on the available data.

Who is the resident TFL Expert?

Dr. C. J. Satchwell and Dr. David Lowe both have extensive world class experience in the management, analysis and forecasting of sophisticated data systems.