Behind every AI trading signal is a machine learning model. But which models are actually used, and what are their strengths and weaknesses? An overview for non-techies.

1. Random Forest: The All-Rounder

Random Forests combine hundreds of simple decision trees into a powerful ensemble. Each tree makes its own prediction, and the majority vote becomes the final signal.

Strengths: Robust against overfitting, works well with many features, requires little hyperparameter tuning. Ideal for classifying market phases (trend/sideways/correction).

Weaknesses: Cannot extrapolate — in completely new market situations, the model fails because it can only interpolate.

2. LSTM (Long Short-Term Memory)

LSTMs are a special type of neural network particularly good at handling time series. They have a memory that can store relevant information over longer periods.

Strengths: Excellent at recognizing long-term dependencies in price data. Can identify seasonal patterns and cyclical trends.

Weaknesses: Slow training, high computational requirements, tendency to overfit on small datasets. Requires large amounts of data for reliable results.

3. XGBoost: The Competition Winner

XGBoost (Extreme Gradient Boosting) has won more ML competitions than any other model. It sequentially builds decision trees, with each new tree correcting its predecessors errors.

Strengths: Extremely performant, built-in regularization against overfitting, fast training. The favorite for tabular financial data.

Weaknesses: Sequential training is hard to parallelize. Sensitive to noise in data.

4. Transformer Models

Originally developed for language processing (GPT, BERT), Transformers are increasingly being applied in financial analysis. They can analyze news sentiment while simultaneously recognizing price patterns.

Strengths: Multimodal analysis — combine text, numbers, and time series in one model. Excellent at processing news flows and earnings calls.

Weaknesses: Enormous computational requirements, millions of parameters, hard to interpret. Nearly impossible for retail investors to train themselves.

5. Reinforcement Learning

Instead of training on historical labels, an RL agent learns through trial and error. It interacts with a simulated market environment and optimizes its strategy through reward (profit) and punishment (loss).

Strengths: Automatically learns the optimal trading strategy including position sizes and timing. Can adapt to changing market conditions.

Weaknesses: Extremely unstable training, requires millions of simulated trades. Often unreliable and hard to reproduce in practice.

Which Model Does the Dieter Investment Club Use?

We employ an ensemble of XGBoost and LSTM, supplemented by Transformer-based sentiment analysis. XGBoost delivers core predictions based on technical and fundamental data, LSTM captures long-term trends, and the Transformer analyzes news flows in real-time.

This multi-model approach compensates for individual model weaknesses and delivers more consistent signals than any single model.