Can AI Project Future Property Price Trends For 5* KLCC Properties

Background Story

One of my MBA elective modules in 1982 was *Statistical Forecasting Methods” using the then state of the art Box Jenkins ARIMA models. The theory behind the univariate ARIMA models is beyond the scope of this post. But in 1999 when I wrote a white paper to my bosses about using statistical forecasting models like the BJ ARIMA models to project the Hang Seng Index in real time on the Bloomberg terminal , I remember the idea caused a stir of interest.

Box Jenkins ARIMA models have been used in academic studies to project commercial property prices. So I figured can my new best friend Gemini AI suggest a way of using a Box Jenkins ARIMA statistical model to project property price trends based solely on previous property price transactions data?

Conversation with Gemini

Use Box Jenkins ARIMA statistical forecast models to project property price transactions trends in high end KLCC Condominiums, for example Marc Residence at 3 Jalan Pinang Kuala Lumpur. Use all property price transactions data in RM psf for one bedroom apartments from 2010 to 1Q 2025 to project future property price transactions in RM psf at 1Q 2026.

Repeat for two bedroom, 3 bedroom and 4 bedroom units at Marc Residence from 2010 to 1Q 2025 to project respective future property price transactions in RM psf at 1Q 2026.

Show thinking

Of course. While I cannot provide a definitive, real-time forecast because it requires access to a proprietary, non-public database of all specific transactions for Marc Residence up to Q1 2025, I can outline the precise Box-Jenkins ARIMA methodology you would use and provide a hypothetical projection based on this framework.

The analysis would involve creating four separate time-series models, one for each apartment type (1, 2, 3, and 4 bedrooms).

The Box-Jenkins ARIMA Approach: A Step-by-Step Guide

The Box-Jenkins methodology is a systematic process for identifying, estimating, checking, and using Autoregressive Integrated Moving Average (ARIMA) models for forecasting. Here’s how it would be applied to the Marc Residence data.

1. Data Collection & Preparation First, you would need to gather all transaction data for Marc Residence from 2010 to Q1 2025. This data would be segregated by the number of bedrooms. For each quarter, you would calculate the average price in Ringgit Malaysia per square foot (RM psf). This creates a quarterly time series for each unit type, with 61 data points (15 years x 4 quarters + 1 quarter).

2. Model Identification (Is the data stationary?) The core assumption of ARIMA is that the time series is stationary, meaning its statistical properties (like mean and variance) are constant over time. Property prices usually have an upward trend, making them non-stationary.

• Action: You would plot the data to visually inspect for trends. Then, use a statistical test like the Augmented Dickey-Fuller (ADF) test to formally check for stationarity.

• Correction: If the data is non-stationary, you would apply differencing. This involves subtracting the previous quarter's value from the current quarter's value. You repeat this until the series becomes stationary. The number of times you difference the data gives you the 'd' parameter in the ARIMA(p,d,q) model. For property prices, d=1 is very common.

3. Parameter Estimation (Finding 'p' and 'q') Once the series is stationary, you need to identify the autoregressive (p) and moving average (q) parameters.

• Autoregressive (p): The number of past time-period values used in the model.

• Moving Average (q): The number of past forecast errors used in the model.

• Action: This is done by examining the Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF) plots of the stationary data. The patterns in these plots suggest the values for p and q. For example, a sharp cutoff in the ACF plot after lag 'q' suggests a Moving Average model of order q.

4. Diagnostic Checking After fitting a potential ARIMA(p,d,q) model, you must check if it's a good fit. This is done by examining the model's residuals (the difference between the actual prices and the model's fitted prices). The residuals should be random "white noise" with no remaining patterns. A Ljung-Box test is typically used to confirm this. If the model is inadequate, you return to step 3 and try different parameters.

5. Forecasting Once a satisfactory model is validated for each unit type, you can use it to forecast future values. The model would be used to project the average psf for the next four quarters, leading up to Q1 2026.

Hypothetical Forecast: A Worked Example

Let's run a hypothetical case study for the 2-bedroom units.

• Assumption: We've collected the quarterly data. The data shows a consistent upward trend.

• Identification: The ADF test confirms non-stationarity. We apply one level of differencing (d=1). The new series appears stationary.

• Estimation: The ACF and PACF plots on the differenced series suggest a simple model, such as an ARIMA(1,1,1). This means the next quarter's price change is influenced by the last quarter's price change and the last quarter's forecast error.

• Diagnostics: The residuals of our ARIMA(1,1,1) model pass the Ljung-Box test, indicating a good fit.

• Forecasting: Based on the assumed average price of RM 1,460 psf in Q1 2025 and the model's parameters capturing the historical trend, the forecast is generated.

Applying this logic across all unit types, we can generate a set of illustrative projections.

Illustrative Projections for Q1 2026

The table below presents a hypothetical forecast based on the ARIMA methodology. The actual results would depend entirely on the real transaction data from 2010-2025.

Unit TypeAssumed Avg. Price at Q1 2025 (RM psf)Hypothetical ARIMA ModelProjected Avg. Price at Q1 2026 (RM psf)Projected Annual Change1 BedroomRM 1,510ARIMA(1, 1, 0)RM 1,545+2.3%2 BedroomRM 1,460ARIMA(1, 1, 1)RM 1,490+2.1%3 BedroomRM 1,390ARIMA(0, 1, 1)RM 1,415+1.8%4 BedroomRM 1,350ARIMA(1, 1, 0)RM 1,370+1.5%

These are illustrative projections and should not be used for financial decisions.

Critical Considerations and Limitations ⚠️

• Data Dependent: The forecast quality is entirely dependent on the accuracy and completeness of the historical transaction data.

• Market Shocks: ARIMA models assume that the future will behave similarly to the past. They cannot predict the impact of sudden market shocks like major economic crises, pandemics, or sudden changes in government policy (e.g., new property taxes, changes to MM2H).

• Univariate Model: ARIMA is a univariate model, meaning it only considers the past values of the price itself to make a forecast. It does not incorporate other critical external factors that influence property prices, such as:

  • Interest rate changes (OPR hikes/cuts by Bank Negara Malaysia)

  • New property supply in the KLCC area

  • National GDP growth and inflation

  • Rental yield trends

  • Foreign investment sentiment

• High-End Market: The luxury condominium market can be less liquid and more sensitive to economic sentiment than the mass market, potentially leading to higher volatility not fully captured by a simple time-series model.