WSWillem Pieter Sijp

Willem Pieter Sijp

*This is an algorithmically generated author record based on publicly available information.

Willem Pieter Sijp has a total of 4 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Must Accurate Automated Home Valuations Be Hard to Explain? Evidence from Cook County

    by Willem Sijp

    In automated house valuation models, accuracy and transparency are desirable, yet they are often thought of as incompatible. Cook County's Assessor replaced township-level linear regressions with a LightGBM gradient boosting model, a leading machine learning valuation approach chosen for its accuracy although harder to explain. We show that no such trade-off may be needed by comparing a a recently developed transparent Bayesian hedonic model against the Assessor's own model, rebuilt or run from its public code and data. In suburban assessments at 1 January 2022 and 2023 and in predictions of 38,177 later sales across all 38 townships, the hedonic model's unadjusted values are more accurate, mainly because their overall level is closer to sale prices, and the model remains as accurate after both models are rescaled to the median ratio of value to sale price. It is also less regressive in two of three comparisons. Unlike the machine learning model, our model provides a predictive interval for each sale price. Measuring by the percentage of house prices within the 80% modelled interval shows that letting the model's error scale depend on the neighbourhood's price level improves calibration from 69% to 77% among the lowest-priced fifth of homes. Transparency costs no accuracy, and the model shows where its values are least certain: among lower-priced homes, where the fairness of assessment is most at stake.

  • Open Access

    Scale and local information in housing markets: a graph-Laplacian AVM benchmarked against machine learning

    by Willem Sijp

    Automated house valuation is often framed as a choice between predictive accuracy and explicit econometric structure. We examine that premise through repeated prospective evaluation of \(13{,}100\) house sales in Inner West Sydney, comparing a graph-Laplacian hedonic model with gradient boosting, random forest and conventional hedonic models. The graph model is the most accurate: it is modestly more accurate than a gradient booster benchmark supplied with previous and nearby sales. The comparisons reveal a broader pattern: accuracy improves when models capture fine local variation in persistent dwelling premia. Restricting the training region improves plain boosting, as earlier work reports, but most of that benefit disappears when previous and neighboring sales are supplied. A penalized coordinate surface improves a standard hedonic regression substantially; the graph model, which replaces that surface with a regularized persistent effect for every dwelling, is more accurate again. Once it is supplied with regionally interacted hedonic coefficients and time indexes, the extent of the fitting region matters little. Local model structure, engineered sales histories and local fitting therefore provide partly substitutable ways of capturing spatial detail. The graph model has the further advantage of additionally estimating local price indexes, persistent dwelling effects and their uncertainty as explicit model quantities.

  • Open Access

    Scale and Local Information in Housing Markets: A Graph-Laplacian AVM Benchmarked Against Machine Learning

    by Willem Sijp

    Automated valuation models are often presented as a trade-off between predictive accuracy and explicit econometric structure. We test this claim in repeated prospective evaluations of 13,100 house sales in Inner West Sydney across 35 quarterly cutoffs from 2015 to 2023. We compare a graph-Laplacian hedonic model with LightGBM gradient boosting, random forest, and conventional hedonic regressions. The graph model achieves the lowest median absolute percentage error: 9.75%, compared with 10.02% for the best LightGBM benchmark, which is supplied with features based on previous and nearby sales. Although modest, the 0.27-percentage-point difference is statistically resolved (95% bootstrap confidence interval: 0.07–0.51). More broadly, the comparisons show that accuracy depends less on whether a model is econometric or machine-learning than on how it captures fine local variation in persistent dwelling premia. Restricting the training region improves plain LightGBM, but most of that benefit disappears once local sales histories are supplied. A penalized coordinate surface substantially improves a conventional hedonic regression; once hedonic coefficients and time indexes vary regionally, the geographic extent of the fit matters little. Local model structure, engineered sales histories, and local fitting are therefore partly substitutable routes to spatial detail. The graph model additionally estimates local price indexes, persistent dwelling effects, and their uncertainty as explicit model quantities. It therefore retains transparent structural outputs without sacrificing point-prediction accuracy, although at substantially greater computational cost than engineered boosting.

  • Open Access

    Graph‐Laplacian modeling of spatiotemporal effects for house price estimation

    by Willem P Sijp, Marc K. Francke

    Many variables involve the modeling of spatial effects, and their dynamics over time. This article presents a linear model in which spatiotemporal random effects are modeled by graph‐Laplacians. A graph‐Laplacian flexibly encodes adjacency in both space and time, in our case not depending on unknown parameters. The graph‐Laplacian can be input for a prior in a Bayesian estimation setup, or used as regularization term in a Ridge regression. A spectral decomposition of the graph‐Laplacian significantly reduces computation time for estimation. As an application, we estimate graph‐Laplacian hedonic pricing and repeat‐sales models on sales prices of Australian residential properties in the period from 1990 to 2024. Bayesian and Ridge regression estimation results are very similar, although the computation time for the Ridge regression is orders of magnitude faster, however at the expense of missing posterior density functions. Our results highlight the advantages of graph‐Laplacians for predicting individual property prices and for producing stable, granular price indexes, also in thin markets.

  • Open Access

    Regional Housing Growth and Relative Valuation Against a National Market Benchmark in Australia

    by Willem Pieter Sijp

    Australian house prices have risen strongly since the mid-1990s, but regional growth has been highly uneven. Using repeat-sales indexes for houses from 1995 to 2024, we estimate a three-factor model in levels that decomposes each regional series into a National Market component and two mean-reverting geographic spreads linked to resource exposure and lifestyle demand. The National Market factor provides a baseline for long-run housing growth, while the spread factors explain cyclical departures from that baseline. We find that regional sensitivities to the National Market factor are reasonably stable over time and reveal that major capital-city markets, especially Melbourne and Brisbane, have participated more strongly in national housing growth, while regional markets have tended to lag. The framework provides a transparent way to summarize long-run regional market exposure and shorter-run cyclical position without estimating fundamental value or forecasting the national housing path.

  • Open Access

    Data‐sparse price indexes by spatio‐temporal regularization and PCA: An application to the Australian housing market

    by Willem P. Sijp, Anastasios Panagiotelis, Marc K. Francke

    We present two novel approaches to overcome the limitations of data‐sparse local house price indexes and combine them into a single model pipeline that is simple, computationally efficient, and interpretable. The first contribution is a new spatio‐temporal regularization of least squares dummy variable models, such as the repeat sales regression used here. This regularization encodes prior knowledge of the proximity of houses in space and their sales in time. It handles missing values in a natural way. The second is nonlocal regularization using truncated principal component analysis (PCA) applied to the resulting national collection of local price indexes. The PCA loadings show that there are important underlying socioeconomic factors that can be leveraged in the construction of Australian market indexes. This PCA reveals important socioeconomic factors, showing that many local markets can be described by a few broad aspects of the national market, consisting of a general trend that contrasts regions influenced by the mining industry with Sydney and Melbourne, and another trend that highlights lifestyle.


  • Open Access

    Exploring the Application of a Mixture Density Network to the Australian Housing Market

    by Willem Pieter Sijp, Anastasios Panagiotelis

    We explore the application of mixture density networks to house price sales data for the entire Australian market and construct price indexes for geographically smaller regions without using many house characteristics, and to track price distributions. We find that the temporal evolution in city house price variance is driven by variations in price behaviour between different locations and not within-location variation: price changes are more pronounced for the expensive segment. Also, individual dwelling valuation via indexation on previous sale prices is more robust to the level and spread of local house sale prices than automated valuations based on house characteristics such as a hedonic function. Furthermore, averaging the estimates from these two approaches yields significantly more accurate home valuations. Our exploration shows a potential usefulness of this type of model for studying time-dependent distribution-based aspects of the market, as well as index-construction in data-poor environments where few house characteristics are available.

  • Open Access

    Stability of Antarctic Bottom Water Formation to Freshwater Fluxes and Implications for Global Climate

    by Jessica Trevena, Willem P. Sijp, Matthew H. England

    The stability of Antarctic Bottom Water (AABW) to freshwater (FW) perturbations is investigated in a coupled climate model of intermediate complexity. It is found that AABW is stable to surface freshwater fluxes greater in volume and rate to those that permanently “shut down” North Atlantic Deep Water (NADW). Although AABW weakens during FW forcing, it fully recovers within 50 yr of termination of FW input. This is due in part to a concurrent deep warming during AABW suppression that acts to eventually destabilize the water column. In addition, the prevailing upwelling of Circumpolar Deep Water and northward Ekman transport across the Antarctic Circumpolar Current, regulated by the subpolar westerly winds, limits the accumulation of FW at high latitudes and provides a mechanism for resalinizing the surface after the FW forcing has ceased. Enhanced sea ice production in the cooler AABW suppressed state also aids in the resalinization of the surface after FW forcing is stopped. Convection then restarts with AABW properties only slightly colder and fresher compared to the unperturbed control climate state. Further experiments with larger FW perturbations and very slow application rates (0.2 Sv/1000 yr) (1 Sv ≡ 106 m3 s−1) confirm the lack of multiple steady states of AABW in the model. This contrasts with the North Atlantic, wherein classical hysteresis behavior is obtained with similar forcing. The climate response to reduced AABW production is also investigated. During peak FW forcing, Antarctic surface sea and air temperatures decrease by a maximum of 2.5° and 2.2°C, respectively. This is of a similar magnitude to the corresponding response in the North Atlantic. Although in the final steady state, the AABW experiment returns to the original control climate, whereas the North Atlantic case transitions to a different steady state characterized by substantial regional cooling (up to 6.0°C surface air temperature).