Having been around for about fifty years, neural networks are hardly new technology. But they appear to be enjoying something of resurgence. The latest return to favour appears to be on the back of a new generation of tools that have made the application of the technology far more straightforward than before. Instead of spending days pre-processing data into a format digestible by the network, many of today’s applications require little more than a spot of judicious pointing and clicking. Though this is of course laudable, there have in fact been some neural net tools that shield the user from underlying complexity on the market for some time.
Market Technologies’ VantagePoint product has been around since the mid ‘90s and consists of a main application that can be coupled with a range of market specific modules for predicting short-term moves in most major futures markets.
Each market module (of which there are twenty-one – covering interest rates, currencies, stock indices, and energies) consists of five separate neural nets, which include data from nine other markets in addition to the target one in their deliberations. These predict respectively: tomorrow’s high, tomorrow’s low, a ‘Neural Index’ which reflects VantagePoint’s view as to whether the market is likely to make a top or bottom in the next few days, a five-day moving average (of closes) two days in the future and a ten-day moving average four days in the future.
VantagePoint is not an automatic trading system, so this output is there to provide decision support rather than outright buy or sell signals.
VantagePoint can display its output in three different formats: a daily report which displays actual and predicted data for a user-definable number of days up to the current date, a history report which is similar to an expanded version of the daily report and also gives the user the option to decide exactly which data is displayed, and a chart window. All three output formats can be displayed in either overlay or displaced mode. In overlay mode, predicted data is overlaid on the date it was generated next to actual data for that same date. In displaced mode predicted data is shifted forward the appropriate number of days, so for example the ten day moving average, which is predicting a value four days into the future will be shifted forwards four days. It is obviously important to be aware of which display mode a report is in, so this is clearly flagged at the top of the report.
The number of ways in which VantagePoint can be used as part of a broader trading strategy are really only limited to the imagination. For example, just the simple prediction of the next day’s high and low would be handy as a timing tool for short-term sellers of option volatility. Alternatively, the predicted five- and ten-day moving averages could provide a basic moving average crossover tool that could be enhanced with information from other tools. (An important point to note here is that because these averages are predicted, they circumvent the persistent problem with conventional moving averages of lag.)
By calculating the difference in ticks between the two predicted averages and the current value for the real moving averages of the same lengths, VantagePoint also effectively provides a couple of handy crossover oscillators (called PTS and PTM). Aggressive traders might care to take tightly-stopped positions based on the shorter of these two momentum tools, while the more conservative might prefer to wait until the longer is pointing in the same direction as well.
Any of these tools could quite feasibly be combined with other proprietary tools that a trader might be using. Alternatively, the full complement of VantagePoint tools could be combined into a standalone trading strategy. For example, for a possible long position one could wait until the PTS and PTM oscillators were in accordance with each other, the Neural Index switched from 0.00 to 1.00, and the predicted high and low for the following day were both above those for the current day.
However, the value of any of the strategies is obviously dependent upon the quality of VantagePoint’s predictions. Three of Market Technologies’ twenty-one VantagePoint market modules were provided for this review – Japanese yen, light crude, and T-bonds. Though the testing undertaken was by no means exhaustive, the results certainly pointed to a very solid performance. For example, the predicted highs and lows for Japanese yen June 99 IMM contract both showed a correlation of more than 0.92 with the actual values achieved. The average prediction error for both highs and lows was marginally over 10% of the average daily range over the life of the contract.
For light crude, the correlation between predicted and actual high/lows was even higher – in excess of 0.99 during the life of the May 99 contract.
The benefits of completely independent neural nets for each predicted value is apparent here – even though the error for the predicted high starts to rise as the daily range expands, the predicted low remains subdued.
The prediction quality is also borne out by client experience, with a number of substantial institutions successfully using VantagePoint in their proprietary trading. Individual clients have also been using the product with positive results. For example, leading Securities America stockbroker Larry Laine started using VantagePoint for his personal trading of the S&P E-mini in early June, and by early September has seen his initial equity increase by a factor of eight. “Previously I was doing little more than tread water with around 50% winning trades,” says Laine. “That win rate has now risen to the 70-80% region. Though the results are partially due to tight money management, a major factor is having VantagePoint to give me a frame of reference – it’s an excellent short-term trading tool.”