A Series on Time Series, Part V: Hybrid Models —Bringing the Best of Both Worlds

In the previous installments of this series, we’ve walked through traditional time series methods like ARIMA, ventured into the world of machine learning, and discussed multivariate approaches using Copulas and Vector Autoregression (VAR). As we reach the fifth part of this series, we will turn our attention to hybrid models—approaches that combine the strengths of traditional statistical methods with cutting-edge machine-learning techniques. These models, designed to capture both linear and non-linear patterns in time series data, provide a powerful toolkit for complex domains like real estate, finance, and proptech.

Why Hybrid Models?

In real-world applications, time series data often contain intricate dynamics that are not easily captured by a single modeling approach. Traditional models like ARIMA and Exponential Smoothing excel at capturing linear trends and seasonality, but they can falter when confronted with non-linear dependencies or structural breaks in the data. On the other hand, machine learning methods such as neural networks and random forests can identify complex patterns but may lack the interpretability and robustness that domain experts value in simpler, statistical models.

Hybrid models seek to bridge this gap by leveraging the strengths of both worlds: the reliability and simplicity of traditional time series methods, and the adaptive, data-driven capabilities of machine learning. By combining these approaches, we can build models that are more accurate, flexible, and robust.

Types of Hybrid Models

There are several ways to combine traditional and machine learning methods in hybrid models. Here are a few of the most common strategies:

1. Residual-based Hybrid Models One of the simplest and most effective approaches is to first apply a traditional time series model, like ARIMA, to capture the linear components of the data. The residuals (i.e., the differences between the predicted and actual values) are then modeled using a machine learning algorithm such as a neural network or decision tree to capture any remaining non-linear patterns. This approach is particularly useful in real estate markets where trends and seasonality may be captured by ARIMA, while price spikes or structural breaks may be better handled by a machine learning model.

2. Parallel Hybrid Models In parallel hybrid models, both traditional and machine learning models are applied simultaneously to the time series data. The final forecast is a weighted combination of the predictions from both models. For example, ARIMA could handle the overall trend, while a machine learning model captures sudden changes or complex relationships. The weights for each model can be learned from the data, ensuring that the model adapts to different regimes in the data.

3. Integrated Hybrid Models Integrated models take hybridization a step further by embedding traditional time series components within machine learning architectures. For instance, recurrent neural networks (RNNs) and long short-term memory (LSTM) networks, which are popular in time series forecasting, can incorporate seasonality and trend components as part of their architecture. This allows the model to learn both the statistical structure of the data and complex temporal dependencies in a unified framework.

Applications of Hybrid Models in Real Estate and PropTech

The real estate market, with its mix of cyclical trends, external shocks, and localized variations, presents a perfect testbed for hybrid models. Here are some key areas where they can offer significant advantages:

1. Property Price Forecasting Hybrid models can improve the accuracy of home price forecasts by capturing both long-term trends (using traditional methods) and short-term price fluctuations (using machine learning). For instance, a traditional ARIMA model could capture the impact of long-term interest rate changes, while a neural network could identify localized spikes or drops in property prices based on additional data like demographic shifts or policy changes.

2. Demand and Supply Forecasting In proptech, understanding the supply-demand balance is crucial for developers and investors. Hybrid models can combine macroeconomic indicators (like mortgage rates and GDP growth) with localized, non-linear factors (such as new zoning laws or infrastructure development) to produce more accurate demand forecasts.

3. Rent Price Dynamics Rent prices tend to exhibit both linear patterns, like seasonal fluctuations, and non-linearities driven by localized economic factors. Hybrid models can capture these dynamics more effectively than either traditional or machine learning models alone.

Challenges in Hybrid Modeling

While hybrid models offer many advantages, they also come with their own set of challenges. These include:

• Model Complexity: Combining two models increases the complexity of both model selection and interpretation. In highly regulated industries like real estate, explainability is often just as important as accuracy, so finding the right balance is critical.
• Overfitting: With more flexibility comes the risk of overfitting the model to noise in the data, particularly when machine learning models are involved. It is essential to use proper validation techniques, such as cross-validation or walk-forward testing, to ensure robust performance.
• Computational Cost: Running multiple models in parallel or sequentially can increase computational costs, especially when large datasets are involved. It’s essential to consider these costs when deciding whether a hybrid approach is worth the investment.

Conclusion

Hybrid models offer a compelling solution for time series forecasting in complex, data-rich environments like real estate and proptech. By combining the strengths of traditional and machine learning approaches, these models can capture both the linear and non-linear patterns present in the data. As data becomes more plentiful and sophisticated, hybrid models will likely become an indispensable tool for analysts and decision-makers alike.

In the next part of this series, we will explore advanced validation and performance metrics, ensuring that the models we build are not only accurate but also reliable and interpretable. 

Franklin Carroll

Franklin Carroll is Kukun's VP of Modelling and Analytics; After finishing his studies at the University of Chicago, he worked at Fannie Mae where he helped build Collateral Underwriter, roll out the Value Verify initiative, and specialized in collateral related and credit related policy.

Published October 14, 2024