GESTIÓN DEL PROYECTO
GESTIÓN TOTAL DEL APOYO VISUAL AL PROYECTO AMEYAL
The goals of this thesis, as stated in chapter 1 were:
1. To investigate the relationships between the intelligent techniques and the applications. This took two forms: the comparing and contrasting o f various techniques for specific problems, and the exploration o f individual techniques across a range of applications.
2. To bring previously proprietary intelligent financial systems knowledge and research to the public domain.
3. To further the field o f intelligent systems through the development of new techniques and the enhancement o f existing methods.
4. To make novel observations, statements and hypotheses about the nature o f the problem domain(s). Most o f the projects have been concerned with trading financial markets.
Each of the projects undertaken in the course of this research had all of these goals, but in addition, also had the requirement o f addressing the business problem. Each o f these projects will now be critically assessed, and conclusions presented.
8.2 Neural Networks for Residual Value Forecasting
8.2.1 Project Summary
The objective o f this project was to use neural networks to forecast the residual value o f new vehicles in 3-4 years time, to an accuracy of £50. A database was available o f the prices of new and second-hand vehicles in a range o f conditions, at monthly intervals for SVi years. A synthetic depreciation series was constructed for proof-of- concept purposes. A pair o f neural network time-series forecasting models were compared to a pair o f linear benchmarks for this data.
8.2.2 Assessment
This project did not meet these objectives and instead has brought some important points to hght on the apphcation o f intelhgent systems. There were several reasons why the system did not operate as desired, and many of these stem directly from the choice o f technique: namely neural network time-series forecasting. These assessments must ultimately stem from the behaviour o f the models under out-sample testing (see Table 8.1):
Table 8.1: Residual value forecasting errors
Technique Variant RM S testing set forecast error (£)
Regression Linear regression 478
Exponential regression 593
Neural network Time-series forecasting 627
1®‘ difference forecasting 3639
It is clear that for this apphcation, for these implementations, that the linear techniques out-perform the more complex non-linear models. This is an important and positive result, as it is a valuable reminder that one does not have to recourse to building large, complex systems to get most o f the information out o f the data.
8.2.3 Project Conclusions
1. Complexity o f technique/model should be appropriate for the complexity o f the problem. The preprocessing that was carried out transformed the problem into one that had httle forecastable non-linearity. The linear models work better because they imphcitly use domain knowledge that is not available to the neural networks. If the synthetic series wasn’t approximately a straight line, regression would not have been used. Regression was used because it was going to w ork fairly well. A neural network must process a reasonable amount o f information before recognising that the system is near linear. If knowledge o f the problem can be used to effect transformations that result in the forecast being trivial (as in this case), then the solution to the problem then becomes one o f “undoing” the preprocessing to yield a forecast.
2. Residual value forecasting is a feasible proposition, but the original project target of an RMS error o f £50 was perhaps somewhat optimistic. This study gives an informed basis from which to make expectations and assessments of the performance of forecasting technologies.
3. Attempting to forecast first differences o f the time-series was ineffective, owing to the sensitivity of the reconstructed depreciation path to systematic errors. An inference from this experimental result is that when httle data is available, model frameworks should be chosen to maximise their robustness, use as much domain knowledge as possible and have as few firee parameters as possible.
4. It is important not to interpret this to mean that it is not possible to use neural networks effectively for this problem, simply that for the final form o f the forecasting, after the preprocessing had been done, the hnear models were more appropriate and consequently were more robust and accurate. If other data was available or a different approach had been taken to modelling depreciation, then it is possible that neural networks would be the most appropriate technology.
8.2.4 Future Work
1. It has been hypothesised in this thesis that the information required to make forecasts sufficiently accurate to meet the £50 RMS error simply was not present
in the data used in this project. A valuable project would be to attempt to enhance the information content available to the machine through the use of macro- economic data. It is possible that neural networks would be useful for this apphcation.
2. An interesting project would be to construct an agent-based Monte-Carlo simulation o f the depreciation o f vehicles, given the forces of supply and demand that WÜ1 operate at each stage o f the vehicle’s life. This is o f particular interest as if
a vehicle has a strong anticipated residual value, it whl have lower hire charges and so probably be in demand. However, at disposal 3 or 4 years later, the market will be flooded with ex-fleet hire vehicles o f that model.
3. Given the existence of a device for forecasting the aggregate depreciation across a vehicle range, a description o f a method o f making individual model residual value forecasts is described in section 3.8.1.
8.3 Genetic Algorithms for Trade Filtering
8.3.1 Project Summary
The aim o f this project was to model expert trading knowledge without recourse to knowledge elicitation exercises. Genetic algorithm rule induction was used on a history o f past chart trades to jSnd rules that capture the knowledge content o f the original trade history and to automate the discovery of new types o f chart patterns.
8.3.2 Assessment
This project was very successful and comprehensively met its brief of capturing, reproducing and extending expert knowledge. This claim can be substantiated by the induction o f a number o f new trading rules on out-sample data that are statistically significant to greater than 95% confidence, and that the distribution of rule Z-scores is positively skewed. This demonstrates that chart patterns must have information content (consistent with [OsPh95]) and that the construction o f useful trading rules can be successfully carried out with genetic algorithm rule induction. As a result, this project meets thesis objectives 1,2 and 4.
1. It is difficult to assess to whether or not the data was complete, honest and reflective of “real-life”, or whether it had been massaged in some way. For instance, it is likely that stationarity would be an issue (i.e. that the expert’s ideas and opinions changed over the two decades that the trading history spans), and possible that some o f the history had been mislaid.
2. The apphcabihty o f the induced trading rules is narrow as it is only the trader who has aU the information necessary to interpret the symbols in a rule. Some o f these symbols such as “Bull” are self explanatory, but most are not, such as “Context: DES”.
3. There are problems with performing repeatable empirical experiments on the financial markets. This will be discussed in greater depth in section 8.6, as this is an issue that is common to three o f the four projects that have been carried out.
8.3.3 Project Conclusions
1. It is apparent that simple rules have higher Z-scores associated with them and appear more robust in general. This is due to two reasons: i) If a rule is simple, broadly speaking it will fire more often than one that has many clauses that must be satisfied. This leads to the rule being tested more thoroughly and hence given a higher significance; ii) If a rule is general it is less likely to be over-fit and consequently simply an artefact of the data. This leads to the more general result that simple trading rules are more likely to be effective than more complex rules. 2. Charts have information content but it is not possible to prove from the available
information whether or not they can be used to beat the market.
3. This experiment has been a successful apphcation o f intelhgent systems, and genetic algorithm rule induction has proved to be an appropriate approach for this problem.
8.3.4 Future Work
The most useful development to this work would be to re-test it on trades that have been entered since the original experiments were carried out. This would be a useful test to carry out as there is no possibihty of data snooping, which is often a serious
criticism o f this type o f study. This would also give the opportunity to assess whether charts have sufficient information content to make appreciable returns and/or beat the market.