Scottish housing market: tax revenue forecasting models – review

7. Summary and going forward

Table 3 provides a summary of the individual model assessments for comparison. The weight assigned to each criterion for model selection will depend on the forecaster's priorities. We provide here a discussion of the broad themes in the literature review and practitioner interviews.

Most of the literature focused on housing prices, with decidedly less attention paid to detailed models for transactions. Little attention is paid to the distribution of prices. Many academic researchers dealt with returns to real estate investment trusts and other indexes that look at investment returns, rather than forecasting average house prices themselves.

The models that were most popular and relevant for forecasting house prices included error-correction models. Particularly, the most influential literature on practitioners in the UK was the work of Geoff Meen (see Section 8: References).

The models that the literature and practitioners suggested would be less relevant are large-scale macroeconometric forecasting models and DSGE models, although these may be suited for policy analysis and other components of budget production.

On identifying turning points and structural breaks, there seem to be few clear forecast solutions, although probit models may contain useful information for whether the current period is a peak or trough, and error-correction models may have some ability to predict large future price declines. Practitioners suggested that identifying a future turning point largely relies on analyst knowledge, experience, and insight.

Studies rarely pointed to clear winners of modelling techniques for accuracy, and the consensus is that conclusions are specific to the region and time period under consideration. For this reason, many practitioners use a suite of models (typically an auxiliary error-correction model that helps inform or is imposed on a macroeconomic model). ^[46] The appropriate models will depend on the resources, technical expertise, and protocol of the Scottish forecasting framework.

Even if a combination of models is not used for the forecast, it can be useful to have a simple and reliable benchmark against which to evaluate alternative forecasts, as emphasized by Stock (2002), among others. Evaluating forecast errors on their own will not indicate the success or failure of a model, unless it can be compared against another feasible alternative.

On communication, the review suggests multivariate regression models are preferred to univariate models, and small systems are preferred to large and complex macro and DSGE models. The IMF's Manual of Fiscal Transparency provides useful guidance on best-practice budget forecasting, emphasizing the importance of linking revenue forecasts to macroeconomic variables, rather than simple trends or autocorrelation approaches. ^[47]

Practitioners almost universally use quarterly data; however, many UK peer-reviewed academic studies use monthly data. The monthly data was chosen largely for comparability with US studies and researchers admit that it restricts analysis to a smaller subset of explanatory variables.

No forecasting approaches are ruled out by data available in Scotland; however the performance of some models may be limited by a relatively short history for estimation. Further, gaps could be identified and improved in the future. To expand available data, a logical avenue to explore would be data for the UK as a whole. However, before the devolution of LBTT, the OBR found that the share of UK SDLT from Scotland varied considerably, averaging between 3.8 to 6.7 per cent ( OBR, 2012a). They also found that the average property price in Scotland was below the UK average. Generally, researchers find that UK-wide data tends to be driven by the London housing market, where prices are likely to be influenced by different factors than prices in Scotland.

Throughout the analysis, it was taken for granted that housing prices were easily defined. There are many alternatives formulas to compute a price index, which vary in their statistical properties, such as the influence of extreme values. Forecasters could experiment with different indexes to see which best translate into revenues.

Although LBTT is a small revenue source, the housing market plays an important role in the economy. Forecasters may wish to estimate future housing demand, rents, and prices for reasons other than revenue, such as macroprudential stability, real estate market stakeholder engagement, or implementing and monitoring social housing policy. Further, the importance of a detailed treatment of balance sheet items such as household capital formation in macro models is being increasingly recognised (for example, see Mian, Rao, and Sufi (2013)). Forecasters may therefore wish to invest significant resources on that basis, and the tax revenue forecast will come out of that process at little marginal cost. If housing transactions and prices are desired only for the revenue forecast, then a simple assumption such as the private sector average may reflect an appropriate share of resources.

Next steps

A broad conclusion from our review is that there are no clear winners of forecast models for the purposes of public budgets. Model selection requires budget officials to set priorities for the model's use (the balance between forecasting and policy analysis) and the forecast's role in the wider budget (particularly in the economic outlook).

Once the model's requirements have been established, model selection will largely be determined by the specific data and circumstances of the Scottish housing market. It is difficult to determine in advance which model will perform best without developing and testing them against one another using out-of-sample forecasts to discover their comparative forecasting properties and relative practical merits. Even if the appropriate model is clear, other models may need to be developed as a benchmark for ongoing forecast evaluation.

Finally, when an appropriate model has been selected, further decisions may need to be made regarding the forecasting of exogenous economic variables and the protocol for integrating (and iterating) the macroeconomic and fiscal models. ^[48]

Table 3: Model comparison

Legend

Rule: Forecasting by technical assumption (rule of thumb, growth accounting model, and external consensus)

Univariate: Univariate time series approaches

Multivariate: Multivariate regression models

VAR: Vector autoregressive models

ECM: Error-correction models

Macro: Large-scale macroeconometric models

DSGE: Dynamic stochastic general equilibrium models

Microsim: Microsimulation models