The MTS (mechanical trading systems) are self explanatory, however, shorting is something we don't endorse, even if the short returns for the model look good. Your certainly don't want to ASSUME that the short returns will be any good at all. If they are not listed, you should calculate them yourself. Don't forget the assumptions for these models -- signals are posted AFTER the market closes, and you need to act on these signals by placing an order for the next days OPEN....that's how the returns are calculated. This WON"T necessarily work for MUTUAL funds, unless you can find a clearing house that will allow you to place order for funds on the open. You should calculate the returns for the NEXT CLOSE, if you wish to see if the MTS might support mutual fund trading on the NEXT CLOSE.
As for graphical models, learn to look for agreement on any given day....a consensus should be reached, so to speak, between the competing forecasts. Use different ranges and different models types. Use the DJI individual stocks to confirm the overall market...use the market internals like AV and DV to confirm the overall market....use available mechanical trading systems to confirm or give early warning for the graphical models.....for details...READ ON!
All the models should be used together to inform trading: look for CONSENSUS with regard to the DIRECTION of the predicted line leading the actual price line, as well as the RELATIONSHIP of the actual price relative to the predicted price (above or below). Look for similar or disparate predictions for related indexes.
Many times a careful inspection of all the models will not give the user any clear idea of where the market is likely to go...this is normal: the idea is to monitor the market predictions and trade ONLY when the models provide information that seems likely to shift the probability distribution in your favor. This amounts to a waiting game, where much of your time is spent waiting for a clear signal.
In general, you can use the following guidelines, but each model has its own characteristics. The best guide is the history graph for the model itself: each model has its own internal relation between predicted and actual.
The 'predicted' line should be viewed BOTH as a point of highest probability in terms of the overall level of the market, and also as an indicator of future direction for the market, especially if the patterns in the prediction line suggest high correlation with the patterns within the actual price, REGARDLESS of the length of the forecast period.
Some models will give very little information about the tiny price oscillations from day to day, but instead show a smooth line, and these tend to be the models that should be viewed as showing the highest probability for the actual price, above which one is increasingly inclined to sell, and below which one is increasingly inclined to buy.
Other models will not give much information about the price level overall as being too high or too low, but rather will show little wiggles that correlate highly with the actual price. These models can be used to inform trades regardless of whether they are generated by a long range or short range model. At times the price forecast line and the actual price will track each other at a fixed distance. In this case the forecast line is not so much a measure of the highest probability of the price level, but forecasts the price movements via the leading part of the forecast line.
At times, models will shift between the two modes described above, or show characteristics of both types of predictive capacity. Inspect each model and decide, based upon its historical performance in the graph, how one needs to interpret the output.
At times a model will stop working, and then start working again quite well. At other times the model will stop working and never recover. When a model stops predicting well it is often left on the site to show users what to expect from models over time as they start to predict less and less accurately.
Note that when a forecast line is showing a relation to the actual price that does not fit its historical relationship, one should probably suspend trading until the forecast has demonstrated that it has returned to good relation with the actual price line. This is especially true for the longer range forecasts, but also applies to the short ones. You can see examples of the models not predicting well for a while, and then starting to predict well again.
This is a tough call, because it is not obvious how large a departure, nor for how long, should be considered normal as opposed to a clear indication of failure in the prediciton. The best guide is a close scrutiny of the historical relationship between the actual and predicted price for that particular model. The other way to test this is retrain the net on several recent temporal periods and see if the accuracy and characteristics are better with the newer models.
All of the above highlights another point about the models: the idea of consensus in the forecasts between models. So, when looking at the different models, look for the predicted changes in direction and pattern, but don't rely on any one model too much.
Regarding periods of EXCESSIVE NEWS or information. The models attempt to find a signal that is constantly present in the market. The more NEWS hits the market, the less signal and the more NOISE is present. Think of the models as finding the markets natual rhythm similar the irregular but not quite random idling of an old car. The idle of this car, for example, is not random and can be predicted, however if we put a person in the drivers seat and tell them to roll some dice and jamn down on the gas, or the brake, whenever they roll 6 or 1, then the motor output will be peppered with randomness. News is sort of like that...the model work much better on markets without too much news. So, the upshot is that when they fail, one needs to wait for the market to return to 'normal', if it seems that excessive NEWS (noise or randomness) was part of the problem. However, a second problem crops up. Since the signal is now corrupted by noise (the actual market has moved in a way that is not idling but more random) then the ANN will try and interpret this...this leads to distortions of the PREDICTED line in the future...so one needs to expect this after a period of excessive news.