Scaling HCNN Inputs and Outputs for Financial Time Series

Summary

The document raises a preprocessing issue in a high-low range forecasting model that uses a hyperbolic tangent state transition. Its outputs are the first nine state-vector elements, so predictions are constrained to the interval from −1 to 1. The forecast variables include price returns, moving averages of highs and lows, Bollinger bands, and intraday open-to-extreme returns for stocks and ETFs.

The central concern is that returns can naturally be represented as signed changes, while price-level indicators such as moving averages and bands may lie outside the network's output range. Those variables therefore need a suitable transformation or normalization if the model is expected to predict them directly. The document poses the question but supplies no answer, scaling procedure, model results, or guidance on fitting transformations without look-ahead bias. It is thus a useful statement of a target-scaling problem rather than a complete modeling method.

Key ideas

• The model's tanh state transition bounds its predicted outputs between −1 and 1.
• Signed returns can be represented as positive or negative fractional changes.
• Price-level features such as moving averages and Bollinger bands may exceed the output range.
• Targets outside the network's range require a scaling or transformation strategy.
• The document does not specify or evaluate a normalization method.

Tags

# Non-overlapping ranges of HCNN' observables and of state transition function


# Non-overlapping ranges of HCNN' observables and of state transition function












In the artcicle Forecasting and Trading the High-Low Range of Stocks and ETFs with Neural Networks HCNN is used for forecasting of nine time-series, namely:

- returns of the lows

- returns of the highs

- five-period exponential moving average of the lows

- five-period exponential moving average of the highs

- five-period lower bollinger band of the closes

- five-period upper bollinger band of the closes

- returns of the open

- same-day return open to low

- same-day return open to high

The proposed state transition function is: $$ s_{t+1}=\tanh(W \cdot s_t) $$ The outputs (predictions) of HCNN are the first nine elements of state vector $s_{t+1}$ so the outputs are within the range $[-1,1]$. The questions are:

- Would it be correct to assume that "returns on ..." are represented as negative and positive deltas, i.e. 10% increase in price is 0.1 while 5% loss is -0.05?

- Five-period exponential moving average of lows will be in general well beyound $[-1,1]$ range of $\tanh$. Shouldn't these values be scaled into the $[-1,1]$ range? How?

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.