Download as pdf or txt
Download as pdf or txt
You are on page 1of 16

Francois Ijaz 1

Predicting Short-Term Movements in the Personal Consumption Expenditures Price Index

(PCEPI) Using the ARMA Model

Bella Francois and Ahmad Ijaz

Department of Finance, Sewanee: The University of the South

FINC 305: Financial Modeling

Professor Jing

December 5, 2022
Francois Ijaz 2

Abstract

This paper explores the short-term fluctuations in the Personal Consumption

Expenditures Price Index (PCEPI) using the Autoregressive Moving Average (ARMA) model.

The PCEPI is a critical indicator of inflation and economic well-being and is utilized by the FED

for monetary policy and economic analysis. Our project uses the ARMA model’s Autoregressive

(AR) and Moving Average (MA) components to capture the momentum and volatility in the

PCEPI's time-series data. For our data analysis, we conducted an in-depth exploration of PCEPI's

historical trends, growth rates, and past volatility. This analysis, coupled with the partial

autocorrelation function (PACF) and the autocorrelation functions (ACF) plots, helped us in our

selection of the ARMA model's order and parameters, ensuring an accurate predictive forecast.

By offering precise predictions of the PCEPI's short-term movements, our research provides

valuable insights into the PCEPI’s forecasting, helping economists and financial analysts in

predicting PCEPI’s movements into the future and what that could mean for the economy in

terms of inflation and economic expansion and contraction.

Keywords: Personal Consumption Price Index (PCEPI), Inflation, ARMA Model


Francois Ijaz 3

Introduction

The suggested research aims to predict the short-term movements in the Personal

Consumption Expenditures Price Index (PCEPI) growth rate using the Autoregressive Moving

Average (ARMA) model to be able to predict changes in inflation. Inflation is an ever-growing

problem in the US and international economies. Additionally, analyzing short-term changes in

inflation can provide an essential context to the overall health of the economy and provide

guidance on investment and business decisions. One metric that helps predict inflation is price

indexes, which, according to the Bureau of Economic Analysis, are “a way of looking beyond

individual price tags to measure overall inflation (or deflation) for a group of goods and services

over time” (Prices & Inflation 2021). Specifically, the Personal Consumption Expenditures Price

Index (PCEPI) “is the Federal Reserve’s preferred measure of inflation” (Personal Consumption

Expenditures: Chain-type Price Index 2023) as it provides the most robust and relevant

information. PCEPI is a measure of the goods and services purchased by “persons,” which is

defined as households and nonprofit institutions serving households (NPISHs) who are residents

of the United States (U.S. Bureau of Economic Analysis, 5-2). The purchases encompass but are

not limited to, durable and non-durable goods, housing, healthcare, transportation, recreation,

education, food, and energy. For a complete summary of the goods and services included in the

PCEPI, according to the Bureau of Economic Analysis, refer to Table 1.1, which is included in

the appendix.

The Fed prefers the PCEPI for several reasons, indicating why it was chosen for this

specific research. First, the PCEPI includes a broader range of goods and services compared to

other inflation measures, such as the Consumer Price Index (CPI), as exemplified by the wide

range of goods and services in Table 1.1. Also, the PCEPI uses a chained-dollar methodology,
Francois Ijaz 4

accounting for consumption pattern changes over time. So, as consumers’ preferences and

spending habits change, the index adjusts to reflect these shifts. The PCEPI also accounts for

more accurate housing and healthcare costs. It uses a measure of imputed rent, which reflects the

cost of living in one's home, instead of using home prices and mortgage payments that indexes

such as the CPI use. It also includes a broader measure of healthcare costs, encompassing not

only out-of-pocket expenses but also the cost of health insurance. The PCEPI also accounts for

imported goods and services, which is crucial for an economy like the United States that engages

in significant international trade. So, this price index can reflect the impact of global economic

conditions on domestic prices. Finally, in terms of the formula accuracy for PCEPI, an economic

commentary published by the Federal Reserve Bank of Cleveland, entitled “The CPI–PCEPI

Inflation Differential: Causes and Prospects,” explains that “PCEPI is based on a Fisher-Ideal

formula, which implicitly allows for the possibility of much more substitution” (Janson et al.,

2020). While the specifics of the formula are too complex to dive into in this report, it is essential

to note that while other indexes, such as CPI, do account for substitution between goods and

services, PCEPI accounts for it much more accurately and thoroughly.

Now that the metric has been explained, it is important to consider why predicting the

short-term fluctuations in PCEPI is significant. As previously mentioned, the changes in PCEPI

are directly related to inflation. Studying changes in inflation allows us to understand and make

informed decisions about several economic and financial areas. First, central banks like the Fed

use inflation indicators like the PCEPI to formulate and adjust monetary policy. Predicting

short-term movements helps policymakers anticipate whether inflation is likely to rise or fall,

informing decisions on interest rates and other policy tools to achieve the Fed’s economic goals,

like full employment and stable prices. However, predicting whether inflation will rise or fall
Francois Ijaz 5

does not just impact the Fed. If consumers anticipate rising prices (inflation), they may adjust

their spending and saving patterns. So, by determining whether PCEPI will increase or decrease,

businesses, policymakers, and individuals can plan for potential changes in consumer behavior.

Also, investors, businesses, and financial markets pay attention to inflation indicators when

making investment decisions. So, predicting short-term movements in the PCEPI helps investors

adjust their portfolios and make strategic decisions based on inflation or deflation expectations.

Finally, and perhaps most importantly, the PCEPI and other economic indicators contribute to the

overall assessment of the economy’s health. Short-term movements indicate whether the

economy is expanding, contracting, or facing potential challenges. This information is valuable

for policymakers and businesses in making informed decisions about economic planning and

strategy. Overall, predicting future fluctuations in PCEPI will allow for a robust understanding of

future economic conditions that will affect consumers, investors, and businesses.

Model

We used the Autoregressive Moving Average (ARMA) model for predicting short-term

movements in the PCEPI. ARMA is a combination of two parts: Autoregressive (AR) and

Moving Average (MA). The AR part forecasts future trends based on historical data, effectively

regressing the variable against its previous values. Specifically, it involves 'lags,’ where each lag

is a previous time point used to predict the current value. For PCEPI, this means utilizing its past

values for future predictions. The MA segment, on the other hand, models the error term

(residuals) as a composite of past errors. These error terms represent the differences between

actual and previously predicted values, assuming that past forecast errors contribute to the

current series value. Using both parts, the ARMA model captures the momentum (through the
Francois Ijaz 6

AR part) and the random fluctuations (through the MA part) in time-series data, making it

suitable for the short-term forecasting of economic indices like the PCEPI.

In the ARMA model, the primary variable is the PCEPI itself. The AR part has lagged

PCEPI values across various time periods, denoted as 't.’ The MA part includes lagged forecast

errors. The model's order is indicated by ARMA(p,q). Here, 'p' refers to the number of lag

observations (AR part), and 'q' signifies the size of the moving average window (MA part).

Selecting appropriate values for p and q involves analyzing the autocorrelation and partial

autocorrelation functions of the PCEPI, ensuring that the model neither overfits nor underfits the

data. We have chosen specific values for p and q based on criteria like the Akaike Information

Criterion (AIC) or Bayesian Information Criterion (BIC), with an in-depth rationale for our

selection presented in the subsequent section.

Data Analysis

The data for PCEPI was drawn from the Federal Reserve Economic Database (FRED), in

which monthly data from 1960 to 2022 was collected, resulting in 745 observations. The

chronological aspect of the data plays a vital role in the time series analysis, allowing the ARMA

model to capture trends over time. It is important to note that PCEPI data does not have units and

instead uses a base year value of 100, meaning that if PCEPI is 110 in a particular month, the

cost of the basket of goods and services has increased by 10% since the base period. So, in order

to make the data more

easily interpretable, the

growth rate was calculated using the log method. As seen above, the log method takes the log of

the PCEPI data in the current period and subtracts the log of the data for the previous period. The

resulting dataset, rawpcepi, then needed to be cleaned to remove the NA value for 1960, as the
Francois Ijaz 7

growth rate cannot be calculated when there is no

previous year and multiplied by 100 to get the

percentage change, as seen in the code to the right.

When the new dataset, entitled pcepi_g, was plotted, a clear picture of the growth rate for

PCEPI was shown. This data, shown below, indicates that the growth rate trends align with

macroeconomic events. For example, the drastic dip in growth rate took place around 2008,

which was when the Great Recession began. This is logical, as the economy was in a downturn,

so consumers were spending less. Specifically, the movements in PCEPI growth are better

visualized when looking at data from the last ten years, seen below, as the more recent
Francois Ijaz 8

macroeconomic trends are even more relevant to the current predictions. For example, there was

a significant dip in growth rate in early 2020 to -0.4%, as consumers were spending less in the

early pandemic, but then a drastic spike after early 2020 back up to around 0.3%. According to

an article written for the Federal Reserve Bank of St. Louis, entitled “What Drives Changes in

Inflation?” the sharp uptick in PCE growth rate since April 2020 is likely due to the increase in

the durable goods inflation rate as “Pandemic-related supply chain disruptions for inputs such as

automobile parts and computer chips drove up the prices for many durable goods” (Chien &

Bennett, 2021). More recently, the PCEPI growth rate dropped from an all-time high in the

middle of 2022 at 0.8% to around 0.2%, the same PCEPI growth rate seen ten years ago in 2013.

When analyzing these more recent results in terms of historical macroeconomic context, the

decline in the PCEPI growth rate is likely attributed to declining inflation since 2022, which

could indicate economic growth in the future.

ARMA Prediction and Results

Now that the current data has been analyzed, the order of the ARMA components must

be determined in order to run the model. Specifically, the SARIMAX function in Python is used

to perform a time series regression on the data using the ARMA model. When running this

function, the order must be specified based on the number of lags and white noise in the order of

(p, 0, q). The autoregression part (AR) is denoted by p, and the moving average (MA) is

detonated by q. To determine p and q, the partial autocorrelation function (PACF) and the

autocorrelation functions (ACF) are used for the respective orders. In order to plot the PACF,

first-order differencing must take place, as our initial PCAF plot had too many positive lags,

which revealed the need to differentiate. The autoregressive term, p, can be determined by taking

the value of the lags that cross the blue significance constraint, shown below. The value of p was
Francois Ijaz 9

chosen as 40, due to the lag at 40 being outside the significance constraint. The same process

was applied for q, using the Autocorrelation function plot. The ACF determines the moving

average component, which is the value of the moving average needed to remove the

autocorrelation from the data. In this case, the value of 35 is outside the significance parameter,

so that value is chosen for q. Also, it is essential to note that when running the code to create the

ACF and PACF plots, a specific number of lags needs to be chosen. The value of 40 was selected

to capture the data’s complexity fully. The PACF and ACF functions determined the order of the

ARMA components for our model, which are given as (40, 0, 35).

After determining the order of the ARMA components, the SARIMAX model can be run,

and the predictions created using the ARMA model can be detailed and analyzed. First, the data

must be split into training and test sets. The data was split into 60 years in the training set and the
Francois Ijaz 10

final two years in the test set, as evidenced by the

code to the right. After splitting the data, the

SARIMAX function is

run on the data, and the

summary table

containing metrics for

analyzing the model is

printed, as evidenced by the code entitled “Fit a SARIMAX model to the data” above. Next, the

training dataset is used to predict future values of the PCEPI growth rate. The steps are set at 48

so that two years into the future can be predicted. This is shown in the code below.

Now that the model has been run and the results have been predicted two years into the

future, we can evaluate the results. Specifically, when plotting the time series data of the PCEPI

growth rate against the predicted data using the test set, it is evident that they follow similar

trends.
Francois Ijaz 11

Also, when examining the recent predicted data more closely, as seen in the graph below, a few

trends are evident. First, none of the predicted values indicate a negative growth rate. This is

likely due to the inability of the model to predict major economic events, such as the COVID-19

pandemic. Also, the forecasted values show significant fluctuations in the growth rate, likely

because the historical data the model was trained on experienced drastic changes in the PCEPI

growth rate. Finally, the forecasted data indicates that while the actual future changes in PCEPI

growth rate, and therefore the exact fluctuations in inflation, are unknown, it is likely, based on

our model, that inflation will increase in the near future but likely drop back down shortly after.

Model Evaluation

To evaluate the performance of our ARMA model, we used several key metrics and

techniques. We used numerous plots, including the Time Series Plot, Growth Rate Plot, and

mainly the SARIMAX model forecast plot. The SARIMAX forecast plot reveals the SARIMAX

model's ability to track the PCEPI's historical movements and project them into the future. The

blue line represents the actual time series data, while the red line depicts the model's forecast. A

visual inspection indicates that the model captures the trend and seasonal patterns well up to the
Francois Ijaz 12

current period. The forecasted values (red line) follow the actual data's direction, suggesting a

good fit.

The Log Likelihood value is relatively high, which typically signifies a better model fit to

the data. The AIC and BIC are considerably lower, implying that the model has a good balance

of fit and parsimony. The Ljung-Box test shows a high p-value (Prob(Q)), indicating no

significant lack of fit in the model. Heteroskedasticity (H) suggests variance in residuals, and the

Prob(H) suggests this variance is significant. This could imply changing volatility in the data,

which may need to be accounted for in the model. The optimization output indicates that the

model took 50 iterations to converge, which is within acceptable limits. The coefficient values

for the AR and MA parts, along with their z-scores and p-values, suggest that many parameters

are not significantly different from zero, questioning the necessity of such a complex model.

The SARIMAX model's forecasts align closely with the historical data, and the model

captures the primary trends and seasonality in the PCEPI. However, the lack of convergence and

the significance of many parameters are concerning and suggest that the model might be

overparameterized. To refine it, we should consider simplifying it by reducing the number of


Francois Ijaz 13

parameters and ensuring convergence to improve forecast reliability. This might involve

re-evaluating the need for all 40 AR and 35 MA parameters, as the p-values suggest that most do

not significantly contribute to the model. Furthermore, addressing the heteroskedasticity and

non-normal distribution of residuals could improve the model's predictive power. While the

SARIMAX model shows promise in forecasting the PCEPI, careful consideration of its

complexity and the statistical nuances revealed in the results is necessary for robust economic

forecasting.

Conclusion

Based upon detailed analysis and prediction of the short-term movements in the PCEPI

growth rate, it is evident that the ARMA model creates predictions that can accurately determine

the future PCEPI growth rate. This is evidenced by the rigorous fitting of the ARMA parameters

using the PACF and ACF plots and the similarity of our predicted data to the actual growth rate

data. In terms of the predictions of the forecasted data, it indicates that inflation will continue to

be volatile, as it has been historically, which can inform investors and consumers to be more

cautious due to an uncertain future. Also, as stated in the results section, the forecasted values

show that inflation will increase in the future, again emphasizing the need for caution and

perhaps for consumers to reconsider their spending habits. However, despite the model’s overall

success, some changes could be made to ensure greater accuracy. As the model evaluation

indicates, the large number of AR and MA parameters may not be necessary, and the prediction

values may be more accurate if the values of p and q were smaller. Also, in terms of further

research, it would be beneficial to run the same model with other price indexes, such as the CPI,

to determine the effectiveness of PCEPI in predicting inflation compared to other measures. It

also could be interesting to plot the real inflation rate data alongside the PCEPI data, both actual
Francois Ijaz 14

and forecasted, to determine the model’s accuracy in predicting the inflation rate and the

economy’s overall health. Also, the various macroeconomic factors that influence the PCEPI

growth rate and inflation that were previously discussed, such as the COVID-19 pandemic, are

not fully encompassed by this model. So, determining a method to include the effect of

macroeconomic factors could also help increase the model’s predictive accuracy. Overall, the

ARMA model was better than expected at predicting the fluctuations in PCEPI growth rate and,

therefore, inflation, as the results indicate, and has proven to be a valuable tool for consumers,

investors, and even the Fed to consider.


Francois Ijaz 15

Appendix

Table 1.1: From U.S. Bureau of Economic Analysis (BEA), “Personal Consumption

Expenditures,” in NIPA Handbook: Concepts and Methods of the U.S. National Income and

Product Accounts, 5–3


Francois Ijaz 16

References

Bureau of Economic Analysis. (2021, August 19). Prices & Inflation. Prices & Inflation | U.S.

Bureau of Economic Analysis (BEA).

https://www.bea.gov/resources/learning-center/what-to-know-prices-inflation

Chien, Y., & Bennett, J. (2021, December 9). What Drives Changes in Inflation?. Saint Louis

Fed Eagle.

https://www.stlouisfed.org/on-the-economy/2021/may/what-drives-changes-inflation

Federal Reserve Bank of St. Louis. (2023, November 30). Personal Consumption Expenditures:

Chain-type Price Index. FRED. https://fred.stlouisfed.org/series/PCEPI/

Janson, W., Verbrugge, R. J., & Binder, C. C. (2020). The CPI–PCEPI Inflation Differential:

Causes and Prospects. Economic Commentary (Federal Reserve Bank of Cleveland), 1–8.

https://doi.org/10.26509/frbc-ec-202006

Kliesen, K. L. (2023, May 30). Scenarios for Inflation in 2023: Base Effects in Action. Saint

Louis Fed Eagle.

https://www.stlouisfed.org/on-the-economy/2023/may/scenarios-inflation-2023-base-effe

cts-action

U.S. Bureau of Economic Analysis (BEA), “Personal Consumption Expenditures,” in NIPA

Handbook: Concepts and Methods of the U.S. National Income and Product Accounts

(Washington, DC: BEA, December 2022, 5–1-5–72).

https://www.bea.gov/resources/methodologies/nipa-handbook/pdf/chapter-05.pdf

You might also like