This paper presents a comparative study that investigates and compares feed-forward neural network (FNN) and adaptive neural fuzzy inference system (ANFIS) on stock prediction using

fundamental financial ratios. Stock trading is a process of buying and selling shares of publicly listed companies on a stock exchange platform. Stock market prediction is an extremely complex and difficult problem because there are simply too many factors and noises

affecting the movement of the price. The main motivation of this research is to develop a feed-forward neural network (FNN) and adaptive neural fuzzy inference system (ANFIS) models to resemble the decision-making process of investment experts based on a stock’s fundamental

financial ratios. Moreover, instead of simply predicting future absolute values of stocks, portfolios that consist of predicted winners or losers are selected and assessed.

FNN is the simplest form of neural network architecture. An FNN consists of at least three layers: an input layer, a hidden layer, and an output layer. 

The ANFIS system consists of rules in IF-THEN form. There are five layers. Layer 1 converts each input to outputs of its membership functions. Layer 2 calculates the firing strength of a rule. Layer 3 normalizes the firing strengths. Layer4 consists of adaptive nodes with a function. Layer 5 sums all incoming signals and delivers a final output.

The methodology includes data preprocessing which are feature dropping, standardization, and filling missing entries and experiments include relative return, portfolio construction, and model validation. The relative return of a stock is the difference between its absolute return and the return of some benchmark. Equal-weight strategy used for constructing portfolios means the

hypothetical investment would be distributed equally across stocks in a portfolio. The result of experiment reveals that “Buy” portfolios and “Sell” portfolios selected by FNN outperform ANFIS. FNN outperforms ANFIS in portfolio selection.

Important points:

• Stock trading is a process of buying and selling shares of publicly listed companies on a stock exchange platform, with millions of investors and traders from all over the world actively involved at any given time when the market is open.
• Many existing studies associated with stock market prediction support the well-known efficient market hypothesis (EMH), according to which the price of a stock at any given time reflects all information available about it and is therefore impossible to predict.
• The main motivation of this research is to develop feed-forward neural network (FNN) and adaptive neural fuzzy inference system (ANFIS) models to resemble the decision-making process of investment experts based on a stock’s fundamental financial ratios. 
• Moreover, instead of simply predicting future absolute values of stocks, portfolios which consist of predicted winners or losers are selected and assessed. The portfolio selection mechanism resembles a more realistic approach to stock investment.
• Feature Dropping :  Features with a high density of missing values consistently across stocks were also dropped.
• Trend Stationarization: Our target variable in this research is quarterly relative returns, while many features from the raw dataset possess a clear global trend with respect to time. These features with global trends could hinder our supervised learning models’ ability to generalize and provide reliable predictions. We therefore take the percentage change between consecutive observations for all features.
• Filling Missing Entries: Mean substitution is used to replace these missing values with the average of their neighboring entries. Mean substitution is based on the assumption that the average of the neighboring observations in a time series is a good guess for a randomly selected observation.
• Standardization: As the scales of features vary dramatically, standardization is applied to all features in order to improve the performance of our prediction models. Fixing the Time Frame: We choose to use data from Q1 1996 to Q4 2017. Stocks with the earliest available observation later than Q1 1996 were dropped from the stock universe.
• Relative Return: The relative return of a stock is the difference between its absolute return and the return of some benchmark. By subtracting overall market performance from the performance of each individual stock, we are able to filter out the factors affecting the broader market.
• Portfolio Construction: In this experiment, we use equal-weight strategy for constructing portfolios. This means the hypothetical investment would be distributed equally across stocks in a portfolio. The portfolios are constructed for every quarter in the validation set during the model validation stage and the test set during the final testing stage.
• Model Validation:FNN and ANFIS models are both validated on the validation set. The FNN uses mean squared error for loss function and Adam algorithm for optimization. For the ANFIS model, subtractive clustering with a cluster influence range of 0.5 is used for defining membership functions and fuzzy rules.
• Both “Buy” portfolios constructed using FNN and ANFIS models outperform the full sample universe of 70 stocks in terms of mean quarterly relative return by a significant margin. On the other hand, both “Sell” portfolios underperform the benchmark.
• FNN outperforms ANFIS in constructing both “Buy” and “Sell” portfolios in terms of mean quarterly relative return.
• The standard deviations of quarterly relative returns for the selected portfolios are higher than that of the benchmark.
• Both FNN and ANFIS are better at identifying losers than identifying winners by a small margin.