Neural Networks for Finance

Let’s start with, what is an Artificial Neural Network?

An Artificial Neural Network or ANN is a software program that mimics the human brain’s ability to classify patterns or to make predictions or decisions based on past experiences. The human brain relies on inputs from the five senses, while the artificial neural network uses inputs from data sets. ANN have three or more layers of neurons. The first layer of neurons, called the input layer, has one neuron for each input to the network. Each neuron in the input layer is connected to every neuron in a hidden area. The hidden area occasionally consists of more than one layer, in which case each neuron in the first hidden layer will be connected to every neuron in the second hidden layer. The last layer in the hidden layer is connected to the outer layer. The strength of the connection between the various neurons varies with the weights allocated to each of the inputs. To start with a neural network, you would typically define a function that would map any value to a value between 0 and 1. This is called a “Sigmoid” function.

This function will be run at every neuron of our network, it is useful in converting probability out of numbers. Then after you add the input, you would like to seed them to make them deterministic as you would be adding random weights. Then you would add the synapses (synapses are the connection between each neuron in one layer to another neuron in another layer). Each synapse has a random weight attached to it. As you train this neural network, the error rate will go down as the allocation of weights which were initially random would autocorrect to a more accurate value.

Selection of appropriate inputs is one of the biggest challenges while designing a neural network. Selecting which inputs will impact NIFTY 50 is a relatively tougher job than selecting inputs for which mortgage application is most likely to default. Because both inputs and target outputs used in time series forecasting are very noisy (the data has a lot of random movement that has nothing to do with the trend). The data used in rating mortgage application are usually more generalized. If there is no income in that month, the person will default. However, in a time series forecast, the NIFTY 50 could go down today and come back up tomorrow and remain up for two months. Is it possible to claim that this fall in NIFTY 50 just noise? Credit Card customer search is another domain where neural networks are used because they require very specific customers to sustain. This ideal customer should spend heavily and not use revolving credit line. Hence, per card revenue will be below per card cost which will result in a lower breakeven point. This is crucial for a bank as the incidental & incremental exceed the revenue resulting in a non-profitable business. Therefore, implementing neural networks to distribute credit cards will maximize profits.

MJ futures claims that they achieved a return of 199.2% over a 2-year period using neural networks. Neural networks can identify trends in data that humans might not notice. For example, Dean Barr and Walter Loick at LBS Capital Management using a neural network with 6 inputs. One of these inputs is the ADX, which indicates the directional movement over the previous 18 days. Two more of these inputs are the current value of the S&P 500, and the net change in the S&P 500 value from 5 days prior. It has also been observed that with a network with 3 hidden layers and 20-40-20 neurons in hidden layers was the optimized network with an accuracy of 94.08% for validation dataset.

One of the most common mistakes that most traders make is taking Standard Deviation as their volatility in all their calculations. However, this is a backwards-looking figure and not an accurate anticipation of the future. To solve this figure an implied volatility could be used or volatility using an Artificial Neural Network. This was done in Shaikh A. Hamid and Abraham Habib’s paper on “Financial Forecasting with Neural Networks”, where they discovered that volatility forecasts as per ANN are more accurate than Barone-Adesi and Whaley (BAW) for pricing American options on futures.

1. They started by selecting appropriate inputs. Technical price data on Treasury bonds was fed along with fundamental data that also effects the market could also be added like fed rates, GDP, money supply, Inflation rate, CPI& Inter-market inputs.
2. Then they moved on to processing the input data by scaling it between 0 to 1 and normalizing the data.
3. Specifying a network is the next step. Therefore, they used a feed forward back propagation network. For the input, they took 11 neurons and twice as many in the second layer (In total there are 3 layers).
4. Then they trained the network with enough historical data.