1.  Prediction Outputs.  The Nets determine their output using a method that is similar to taking a weighted average of all the outputs of known patterns with similar inputs. Therefore, mathematically, the output of a Prediction Net can never be larger than the largest Actual value in the known patterns. So if you are predicting prices in a rising market, and prices have never been above 50, then no prediction will ever be above 50 either. You can’t average a bunch of prices below 50 and get a value above 50.

Similarly, no output can ever be below the lowest output of the known patterns. You can’t average a bunch of prices above 10 and get an average below 10.

In this respect, the Adaptive Nets are quite different than the Turboprop 2 neural net algorithm in the Trader, which extrapolates very well. Prediction Nets don’t extrapolate at all. Therefore, don’t predict prices unless you are positive that your known patterns (in the lookback) are all around the same price levels. Predict price changes (momentum) or percent price changes.

2.  Suppose you use a Classifier Net to give you a "Buy" and "Sell" probability. You may be tempted to consider output probabilities as the probabilities you should actually buy or sell. That is not what the probabilities are; they are the probabilities that the pattern being evaluated match known patterns which were classified as "Buy" or "Sell".

That is a BIG difference, although it may seem subtle at first. No model (Net or any other kind of model) can give you a probability that you should buy or sell. All they can do is give you a measure of how the current condition matches past conditions in which a buy or sell was appropriate. The markets could change, or your inputs may not be as predictive as you think they are.

Nets (wisely) sometimes do not give you any output at all. The Net is telling you that it has no basis for making a prediction. It is saying that none of the known patterns are similar enough (given the contribution factors) to be used for a prediction. This is not a bug; it is a protection for you! Isn’t it better not to have an output than an incorrect one?  Since these Adaptive Nets may not produce an output, you may get many consecutive bars where there is no output. If you are using the indicators as inputs to the regular nets, or in a trading strategy, you may get a message that you have "insufficient data" because there are so few bars. This can often happen when optimizing if you allow large contribution factors.

3.  To avoid missing outputs, you can try two things: (1) increase the number of training patterns (the look back period) and/or (2) reduce some contribution factors. Either of those things will increase the likelihood that similar patterns will be found in the known (training) set.

4.  It should be noted here that inputs to Ward Systems Group Adaptive Nets, unlike inputs into other Ward Systems Group neural nets, are not scaled to the same number range. This means that inputs with larger numbers (like volumes, say) may swamp inputs with lower numbers (like percent price change) unless you assign lower contribution numbers (usually less than 1) to the inputs with large number ranges. Don’t be afraid to use very small numbers like .001.