A Series on Time Series, Part I: Why Forecast?
“The only function of economic forecasting is to make astrology look respectable.” — John Kenneth Galbraith
The above quote shows that it is possible to become a famous economist and still be quite confused about the value of forecasting. Economic forecasting, a complex discipline that combines economic theory, data analysis, and statistics, plays a critical role in the decision-making processes of investors, business leaders, and policymakers. This essay will explain why forecasts are necessary, despite some inevitable inaccuracy, and delve into what makes creating them so challenging. We will pay particular attention to the difficulties real estate economists face when producing them. Subsequent essays will discuss how to create forecasts using traditional statistics and machine learning.
Why Bother Forecasting?
Forecasting is challenging, and almost every forecast will be “wrong” in the sense that the actual value will rarely match the forecasted one exactly. This raises the question: Why bother forecasting at all? Despite its imperfections, forecasting is essential. Here are just some of the reasons to produce a forecast.
Setting Expectations
A forecast helps set an organization’s expectations, allowing it to adjust its behavior accordingly. For example, if a time series analysis forecasts certain sales figures, the company can prepare to make adjustments if sales fall below projections; similarly, they can choose to stick with what’s working if sales come in above projections. Establishing a “flight plan” helps the organization communicate and refine its strategy, avoiding counterproductive actions based on differing internal perspectives. In essence, a forecast helps clarify what is “good,” “okay” and “bad” by defining it as “above projection,” “at projection” and “below projection.” Of course, the market will provide companies with feedback in one way or another, even through the harsh measures of bankruptcy. Still, it is better to have a working definition before a company reaches that point.
Stating Assumptions Explicitly
The practice of forecasting forces decision-makers and modelers to explicitly state the assumptions underlying their forecasts. When a forecast is incorrect, these assumptions can be examined to identify which ones were flawed. This provides a framework for self-assessment. For example, if a forecast assumes stable interest rates, a subsequent rate hike would reveal the impact of this assumption on the forecast’s accuracy. The business might conclude thereafter that it needs to hedge against interest rate risks. Of course, this is a simple example, but in the real world, the set of assumptions involved can become quite long and complex. If the model accurately predicts outcomes with the subsequently observed real-world inputs, it can help identify which assumptions were most flawed. If it doesn’t work now but did work in the past, on the other hand, this could indicate something occurred that your planning completely missed, likely because it was not a factor in the past. Listing assumptions in advance speeds up the process of self-criticism and keeps companies nimble.
Confidence Intervals
A proper forecast should include a confidence interval. While the point estimate may often be wrong, the confidence interval helps planners focus on the most likely outcomes. Moreover, confidence intervals provide certainty regarding the general direction of trends. For instance, a 95% confidence interval that contains only upward trends can be useful to planners even if the exact projection is wrong: They know that they do not have to put as many resources into planning for, let’s say, a reduction in sales as they might otherwise. This is why confidence intervals are crucial in academic research, often more so than p-values. While significance is important, a confidence interval crossing zero—meaning you aren’t sure whether the effect you are testing for is positive or negative—means you can throw out your finding regardless of its being statistically significant. A quality forecast, combined with a confidence interval, offers qualitative certainty about whether things are likely to improve or deteriorate. Even if the point estimate is wrong, the overall direction the forecast is pointing in will most likely be right.
Forecasts vs. Guesses
A forecast is generally more useful than a guess, even a good guess. Casinos make money by being slightly more right on average than the customers they are playing against: Indeed, the house edge is a mere half a percent in Blackjack if the player uses basic strategy, becoming as high as 2% if the player is more “creative.” In craps, the house edge varies from 1.4% to 2%. These small margins built the city of Las Vegas. Companies whose planning is based on slightly more accurate estimates can make more money in the long run. This common-sense fact is sometimes forgotten due to the way grading in school emphasizes exact correctness. In the real world, what matters is being closer to the truth than others and making slightly better decisions than your competition over and over again (while avoiding catastrophe, of course, so your long-run edge can have time work).
The Complexities of Home Price Forecasting
Home price forecasting has become more difficult recently. Of course, there are certain difficulties forecasters always face, but there are others particular to this market.
Difficulties Forecasters Always Face
Data Heterogeneity and Volume
Real estate data is highly diverse, encompassing everything from macroeconomic indicators to micro-level details like home amenities. The sheer volume and variety of data can be overwhelming, necessitating sophisticated tools for analysis and interpretation. For instance, luxury homes might exhibit higher price variability compared to more standardized, lower-end homes. This variability in prices contributes to heteroskedasticity (Home), a word that refers to the presence of variables that are correlated with the size of your error terms. In a later section, we will discuss methods that modelers can use to address heteroskedasticity and why it can be such a problem.
Market Volatility
The housing market is inherently volatile, influenced by predictable elements such as interest rates and unpredictable factors like political events or natural disasters. This volatility makes accurate forecasting particularly challenging.
Non-Linearity
The relationship between predictors and home prices is often non-linear. For instance, doubling the income level in a region might more than double the average home price due to increased market demand, making linear models insufficient.
Spatial and Temporal Correlations
Real estate markets exhibit significant spatial and temporal correlations; prices in one location influence prices in nearby areas, and past prices can indicate future trends. Capturing these dependencies requires advanced modeling techniques. However, the more these relationships are captured, the higher the risk of overfitting the model to the data. Overfitting is a perennial risk, especially in machine learning models.
Policy Effects
Housing prices are affected by policy decisions, which are often made by a small group of powerful individuals. While collective behavior has some predictability, the decision-making of elites is not. For example, decisions by the Federal Reserve on interest rates can dramatically alter the housing outlook, as seen recently.
Difficulties Created by the Unusual Market We Are In
Low Volume
Transaction volume is currently very low because few homeowners want to switch from their low-interest loans to much higher-interest loans: This phenomenon has been named “The Locked-In Effect.” This lack of data can hinder the accuracy of machine learning methods, which are data intensive. Moreover, repeat transaction indices are more error-prone when volume goes down leading to heteroskedasticity: Heteroskedasticity is just statistical speak for having error sizes that are correlated with your variables. Statistical models can suffer from a host of problems when this occurs.
Interest Rate Increases Not Seen in Recent Data
It can be difficult to incorporate significant changes like recent interest rate increases into models when there are no similar changes in the dataset. This situation forces reliance on a priori theorizing, which is generally less accurate than empirical approaches.
Forecasting is a critical, albeit complex, task–f-esp. when dealing with the real estate market. Despite its challenges and inherent imperfections, it provides valuable insights by setting expectations and facilitating analysis. The current market conditions introduce additional challenges, making sophisticated analysis and cautious interpretation of forecasts even more crucial. In the next entry, we will discuss the methods forecasters use starting with traditional statistical methods and then turning to machine learning-based ones.