Potential research topics for joint research: Forecasting oil prices with forecast combination methods Dean Fantazzini Koper, 26/11/2014
Overview of the Presentation Introduction Dean Fantazzini 2
Overview of the Presentation Introduction Some more details Dean Fantazzini 2-a
Overview of the Presentation Introduction Some more details Additional research projects Dean Fantazzini 2-b
Introduction The real price of oil is an important variable for macroeconomic forecasting and risk management, and many sectors of the economy depend directly on forecasts of the price of oil to develop their business strategies. Moreover, this forecast is of key relevance for long term investments, carbon emissions regulations, climate change modelling, energy policy regulations, energy system planning. See e.g., Alquist et al. (2013), Hamilton (2009, 2011, 2013), Kilian (2008, 2009, 2010), Fantazzini et al. (2011, 2014) and references therein. Dean Fantazzini 3
Introduction Baumeister and Kilian (2014) established that an equal-weighted combination of four recently proposed oil price forecasting models is systematically more accurate than the no-change forecast as well as forecast combinations based on recursive or rolling inverse MSPE weights. Baumeister, Kilian and. Lee (2014) extended the previous work and compared three approaches to generating real-time oil price forecasts: one assigns equal weight to all forecasting models under consideration. Another allows the subset of models selected for the forecast combination to vary by horizons according to its ability to reduce the MSPE. a third approach involves selecting for each horizon the model with the lowest recursive MSPE. Of these approaches, they found that only the first two can be recommended Dean Fantazzini 4
Introduction On the other hand, Fantazzini and Fomichev (2014) proposed new models to forecast the real price of oil on the basis of macroeconomic indicators and Google search data. Moreover, Kristoufek (2013) proposed a novel approach to portfolio diversification using the information of searched items on Google Trends, where popular stocks are penalized by assigning them lower portfolio weights and less popular stocks are brought forward. Idea: merge forecast combination methods with Google-based approaches to improve short, medium term and possibly long-term forecasting of the oil price. Dean Fantazzini 5
The forecasting models by Baumeister and Kilian (2014) and Baumeister, Kilian and Lee (2014): Reduced-form VAR model: B(L)Y t = ν + u t where Y t = [ prod, rea t, rt oil, inv t ] refers to a vector including the percent change in global crude oil production, a measure of global real economic activity, the log of the U.S. refiners acquisition cost for crude oil imports deflated by the log of the U.S. CPI, and the change in global crude oil inventories, ν denotes the intercept, B(L) is the autoregressive lag order polynomial of order 12 and u t a white noise innovation. ˆR oil t+h t = exp(ˆr oil,v AR t+h t ) Dean Fantazzini 6
Forecasts based on the price of non-oil industrial raw materials ˆR oil t+h t = Rt oil h,industrial raw materials [1 + πt E t (π h t+h)] where R oil t denotes the current level of the real price of oil, h,industrial raw materials πt stands for the percent change of the Commodity Research Bureau (CRB) index of the spot price of industrial raw materials (other than oil) over the preceding h months. The term E t (π h t+h) is the expected U.S. inflation rate over the next h periods. In practice, this expectation is proxied by recursively constructed averages of past U.S. CPI inflation data Dean Fantazzini 7
Forecasts based on oil futures prices ˆR oil t+h t = Rt oil [1 + ft h s t E t (πt+h)] h where f h t is the log of the current WTI oil futures price for maturity h, s t is the log of the corresponding WTI spot price, and E t (π h t+h) is again the expected inflation rate over the next h periods Dean Fantazzini 8
Time-varying parameter model of the gasoline and heating oil spreads Many market practitioners believe that rising spreads between the price of refined products (such as gasoline or heating oil) and the price of crude oil signal upward pressures on the price of crude oil. However, forecasts based on product spreads is unstable over time: 1) One concern is that the price of crude oil is likely to be determined by the refined product in highest demand and that product has changed over time. 2) Another concern is that crude oil supply shocks, local capacity constraints in refining, changes in environmental regulations or other market turmoil may all temporarily undermine the predictive power of product spreads Dean Fantazzini 9
time-varying regression model! s t+h t = β 1t [s gas t s t ] + β 2t [s heating t s t ] + ε t+h where s gas t s heating t is the log of the nominal U.S. spot price of gasoline ε t+h NID(0, σ 2 ) is the log of the nominal U.S. spot price of heating oil the time-varying coefficients θ t = [β 1t β 1t ] evolve according to a random walk as θ t = θ t 1 + ξ t, and ξ t is independent Gaussian white noise with variance Q. oil ˆR t+h t = Rt oil exp{β 1t [s gas t s t ] + β 2t [s heating t s t ] E t (π h t+h)} Dean Fantazzini 10
Forecasts based on U.S. crude oil inventories where ˆR oil t+h t = R oil t (1 + ˆβ inv h t ) inv h t denotes the percent change in U.S. crude oil inventories over the preceding h months and ˆβ is obtained by regressing cumulative percent changes in the real price of oil on the lagged cumulative percent change in U.S. inventories without intercept (the latter restriction improves the accuracy) Dean Fantazzini 11
Fantazzini and Fomichev (2014) generalized Kilian and Murphy (2014) in two ways: 1) expanded the original set of variables (global crude oil production, real activity measure, global above-ground crude oil inventories, real price of oil) including also Google search data; 2) use multivariate cointegrated models (VECM) including both Google data and macroeconomic aggregates and performed out-of-sample forecasts ranging from 1 month to 24 months ahead, comparing almost 50 alternative model specifications. Dean Fantazzini 12
Google data represents how many web searches were performed for a particular keyword (or keywords) in a given week and in a given geographical area, relative to the total number of web searches in the same week and area. This index is then rescaled by Google between 0 and 100 dividing it by its largest value and multiplying the result by 100. More specifically, Fantazzini and Fomichev (2014) used the following online queries: oil+wti+brent : this is a very general query, intended to collect all searches oriented to short term news and events related to the oil industry and financial futures. Dean Fantazzini 13
oil supply : this is a more specific query, intended to collect all searches related to the oil supply in a medium and long term horizon: an increase of searches for oil supply may indicate both a quest for additional oil supplies (which would be positive for oil prices), as well as a quest for information about newly discovered and developed oil supplies, like US shale oil (which would put negative pressure on oil prices). Dean Fantazzini 14
oil demand : this is a specific query intended to collect all searches related to the oil demand in a medium and long term horizon: increase of Google queries for oil demand may be symptomatic of an higher demand for oil, (like during the years 2004-2008); however, oil sellers may also investigate the world situation of oil demand to fine-tune their production. Dean Fantazzini 15
Figure 1: Google search data for oil supply and oil demand. January 2004 - November 2014 Dean Fantazzini 16
Figure 2: Google search data for oil supply and oil demand. January 2013 - November 2014 Dean Fantazzini 17
Additional research projects Modelling the dynamics of Slovenian gasoline and international oil prices We want to verify whether the transmission mechanism of positive and negative changes in the price of crude oil to the price of gasoline in Slovenia is asymmetric. The study of the existence of asymmetry in gasoline prices transmissions can be useful given the potential presence of collusion in the industry, assuming that such asymmetries may be caused by the exercise of market power. On the other hand, a study of the existence of asymmetry in gasoline prices can help to create a list of methods able to provide empirical evidence on the existence of anti-competitive conduct in the gasoline industry. Dean Fantazzini 18
Additional research projects Asymmetric error correction models can be employed to study the gasoline-oil dynamics. Given the limited literature about the Slovenian gasoline market, this project aims to fill this gap: this study can be performed at the national level (using national average prices for gasoline, diesel and oil), or at the regional level, using a disaggregated data set. Dean Fantazzini 19