Timing Solution research project #1

NN model to reveal turning points

Outputs - what we plan to forecast. De trended zigzag and CIT wave indicators. We try to make a prediction for upcoming turning points; at least we try to build high probability turning points zones. The best way to calculate turning points is to use a zigzag indicator. Here is how this indicator looks for S&P500 index:

And there is a problem with a classical zigzag: NN works with DETRENDED index while classical zigzag shows some trend (together with turning points). So, instead of a classical zigzag we should use a detrended zigzag. This is how it looks together with the classical zigzag:

Tops here correspond to top turning points, bottoms - to bottom turning points. In terms of forecast: when this indicator is high, we are expecting a top turning point, and low is for bottom turning points.

We can also apply a non directional zigzag index - CIT wave. Tops here correspond to turning points, whatever they are - tops or bottoms. In other words, when this indicator is high, we have (or expect in the case of forecast) top or bottom turning points. When indicator is low, we are in the middle of a trend, no turning point here. This is how this indicator looks:

Curvature - make forecast sharper. In NN you can specify any zigzag based indicator here:

also I recommend to use a curvature indicator, it emphasizes turning points. Try to set it to 3-5:

This is how it looks:

Summary: As an output, I recommend to use de trended zigzag (directional turning points indicator) or CIT wave (non directional).

Inputs - what the forecast is based on... We can apply different models that reflect different theories. For example, we can build a model based on astro events such as aspects, midpoints, planetary positions etc.

Let us create Harmonic aspect confluence model (HACM). This model assumes that 1H aspect affects the stock market behavior if it is confirmed by (2H) or/and (4H) aspects.

The NN does all the job of finding if this assumption is a true one. The only thing for us to do is to define 1H, 2H and 4H aspects. For 1H we will analyze conjunctions only, for 2H and 4H aspects five major aspects. With Timing Solution, we can play with that model by varying aspects set, involved planets, orb harmonics.

 Run the model editor:

Click "Selection" button:

Firstly create a list of 1H conjunctions aspects:

2H major planets:

4H major aspects:

After that save this model into a file. Let it be HACM.hyp file:

Creating Neural Network: as an output, let us use the detrended zigzag, set the curvature parameter as 3, 4, or 5; as an input, we download the file with the previously created model (in this case it is the file HACM.hyp).

Non linearity, NN hidden layers

Our aspect confluence model contains some non linear factor. We assume that 1H aspect affects the stock market behavior if it is confirmed by 2H and/or 4H aspects. Like Sun conj. Jupiter is acting if we have at the same time some Sun aspect in 2H chart and/or 4H chart. It means that all these aspects should meet together, otherwise we cannot assume that this aspect has any effect on the stock market.

This is a non linear model; the aspect in consideration has some conditions and it cannot be isolated (that is why we cannot use here linear models like Bradley aspect model). NN works well with non linear models, this technology has been developed for that. The question is: what is the degree/grade of non linearity?

Look at this table:

This is a degree of complexity of our system. In NN we set this parameter using the amount of hidden layers parameter, here it is:

The degree of non linearity here is 2. Be warned, to train NN with 2 hidden neurons need more time.

 Pay attention, we do not care how these aspects work in combination with other aspects. The only thing we do is to define aspects themselves, all nonlinear connections between aspects (like some aspect works in combination with a certain aspect only, otherwise it does not work). How they work, the NN finds itself during the training process. This information regarding nonlinear connections is stored in hidden neurons.

Verification, out of sample data. I recommend to put aside some price history to verify our model, to see how it is able to forecast the future.

I set LBC to the end of 2009, so we have almost 9 years out of sample data, this interval covers 16 turning points:

In other words we will see how our HACM model actually forecasts last 16 turning points.

Training.  Run training button and wait... Time to time select some piece of price history on IN SAMPLE (before LBC) data and watch how this projection line coincides with the detrended oscilator index:

For well-trained NN they should coincide very well.

At 150K training steps I have got this:

At 210K steps:

At 300K steps:

At 1.4M steps:

You see, the projection line is too choppy. To make it smoother I've increased the orb for our aspects, I set it to 10 degrees (before it was 5 degrees).

At 200K step I have got this:

At 1M step:

At 1.4M step. It looks better. Isn't it?

At 3M steps, very good:

I've zoomed some price interval to see details:

Moment of truth. Now we need to know how this model really works. In order to do that we need to see how this forecast works after LBC, on out of sample interval.

I've selected a piece of price history on out of sample interval and got this:

Another interval:

And one more interval:

I don't know, it seems to me this model does not work well. We need to do more research.

Our goal is to get an acceptable forecast on the out of sample interval.

What can be researched for this financial instrument? We may try different models - like FAM, dynamic model, etc. We may take for a training interval a  different part of the available price history (which is 1950 - 2009, as shown above), like only last 8 years of it (it will be the years 2001 - 2009). To set training interval to 8 years (2000 bars for EOD chart) do this:

And check all models for that interval. So many possibilities... Does it sound like a plan?