Technical Analysis of Stock Prices: Inherent Flaws and Proposed Model

 

Data used: Boeing Co. Data as of 17 April 2025

It is true that short-term modeling and predictions of stock prices using just price data is valid and useful. We do not need fundamental data as input variables for short-term predictions. But there are inherent flaws in traditional Technical Analysis (TA). The inherent flaws of traditional technical analysis indicators, such as RSI, MACD, Bollinger Bands, all assume that the relationship between market variables is linear and that data distributions are Gaussian (Normal). But it is well-known that financial markets exhibit non-linear dynamic characteristics with distributions that are not Normal i.e. have more than 1 peak, are highly skewed and have long fat tails (kurtosis). And that the relationship between market variables is highly non-linear. 

However, these short-term linear relationships can be modeled with Linear Regression . The Table below shows that Linear Regression, particularly its Boosted version produces the smallest Mean Average Deviation. No need for fancy Neural Networks. 

Here are a few examples:

Here is a raw histogram of Boeing's stock price: 

Note how far it deviates from the assumed Gaussian Distribution of TA.

Proposed improved Technical Analysis model with probabilistic predictions:

• We first run all the models in our ‘arsenal’ of tools to find the combination of models that would yield the best performance for a specific stock. 
• AI models like Neural Nets, Decision Trees and Random Forests handle non-linearity easily through their activation functions, nodes and hidden layers of neurons;
• Econometric models such as ARIMA and Holt-Winters Smoothing help to distinguish between noise and true signals;
• Decision Trees and Random Forests of Decision Trees are used for data where Neural Networks tend to overfit;
• Ensembles harness the output of multiple copies for a consensus that reduce the margin of error;
• Statistical Bagging (random sampling with replacement) and Boosting (enhanced gradient descent) ensembles increase model performance; 
• Monte Carlo simulations test the model by running 1000 trials and establishing the Probability Density Functions (PDF);
• Results are fitted to unbounded MetaLog PDFs which are best suited to are financial markets data to establish the 5% and 95% Confidence Levels for Support and Resistance prices.

A note on ARIMA

ARIMA stands for Autoregressive Integrated Moving Average. It is a technique used by Economists to forecast economic time series, which have some common characteristics with financial markets data. For example, current values are related to past values (in the sense of past values having an impact on current values), and each day previous to the current day is 1 lag. The impact of the lag diminishes the further away they are from the current day. ARIMA can also be used for economic time series that are seasonal such as the demand for Natural Gas or electricity for heating in the Winter months.

But the main reason why ARIMA is important in modeling financial times series is that using it we can glean insights from the data to build our models with more sophisticated technologies such as Artificial Intelligence. In the paragraphs below we explain the various concepts used in ARIMA.

Autocorrelation is the ‘memory’ of the time series. Past values affect the current value.

 

Here are some of graphics to illustrate the procedure outline above with notes below them for explanation.

Post-ARIMA histogram is a big improvment from the raw histogram above it in reducing the skewness of the data. However, note that the outliers still persist both on the left hand side (lows) and right-hand-side.  The ARIMA is not fully effective. The Studentized Residuals chart below show the sudden volatility spike at the end of the series due to the current US-China Trade War and cancelling of Boeing aircraft order by Chinese airlines.

 

Next we do a 1000-trials of Monte Carlo Simulation with the fitted ARIMA series which, when fitted to an unbounded Metalog Distribution looks like this:

So you see it is much better now. Note on Metalog Distributions: One of the unique characteristics of our models is the use of a recent innovation in Statistics- MetaLog Distributions.  Financial markets data are often complex, noisy, and exhibit characteristics such as multi-modality and skewness that are not well captured by traditional Normal (Gaussian) distributions.

The MetaLog Distribution is a flexible distribution that can model a wide range of data types, including those with skewness, heavy tails, and other irregularities. It is known for its ability to fit both bounded and unbounded data with varying degrees of smoothness and flexibility.

Unlike traditional distributions that require assumptions about shape parameters, the MetaLog Distribution adapts to the empirical data, making it highly versatile for various applications.

Metalog Distributions also produce a table that enables you to determine the Support (5% quantile) and Resistance price (95% quantile) for short term trading:

 Now our model will output the predicted autoregressive trend of Boeing for 1 to 20 steps ahead:

As we can see, Boeing is impacted by the ongoing US-China Trade War, and China's cancelling of all Boeing orders by its airlines.