Recent academic papers illustrate how macroeconomic data support predictions of energy market flows and prices. Valid macro indicators include shipping costs, industrial production measures, non-energy industrial commodity prices, transportation data, weather data, financial conditions indices, and geopolitical uncertainty measures. Good practices include a focus on “small” models and a reduction of the dimensionality of large datasets. Forecasts can extend to predictions of the entire probability distribution of prices and – hence – can be used to assess the probability of breakouts from price ranges.

The below is a summary of recent literature by Robert Brown (Economist at Adrian Lee & Partners). The underlying papers are listed at the end of the post.

The post ties in with this site’s summary of systematic value and macro trends.

### In a nutshell

There is ample evidence that data on macro conditions provide valuable information for forecasting both energy prices and energy demand and can improve forecasts significantly compared to both the random walk and univariate models. The following findings are of particular importance:

- Proxies for global economic activity, constructed through dimensionality reduction techniques such as principal components, have produced substantial improvements in forecasting performance while mitigating dimensionality concerns.
- “Smaller” models with fewer parameters perform better than more heavily parameterized frameworks in many settings, and exploitation of shrinkage in larger forecasting models can add significant reductions to forecast errors (Baumeister, Korobilis and Lee (2020)).
- Allowing for time variation in volatilities can substantially improve forecasting performance, especially at longer horizons.

### Prominent examples of macro forecasting models

Ferrari, Ravazzolo and Vespignani (2019) utilize an updated version of the Mohaddes and Raissi (2018) “Global VAR” dataset containing around 200 macroeconomic time series for the 33 largest economies to produce forecasts for the price of coal, gas and oil at a quarterly frequency for up to four quarters ahead. Their dataset includes information on economic activity, interest rates, inflation and financial conditions. They employ baseline vector autoregressive (VAR) and autoregressive models, as well as VAR and autoregressive frameworks which are augmented by factors extracted from the global VAR dataset.

Out-of-sample forecasting results give mixed conclusions. The baseline autoregressive and vector autoregressive models consistently generate larger mean squared forecasting errors compared to the random walk, for all target commodities and all forecasting horizons. Encouragingly, __factor-augmented models exhibit forecasting improvements compared to the random walk__ when forecasting oil and coal one quarter ahead, and across all forecast horizons for gas. In general, the authors find that factor-augmented autoregressive models tend to outperform more heavily parameterized factor-augmented VARs.

Alquist, Bhattarai and Coibon (2018) provide additional evidence, as well as a theoretical justification, for the utility of macroeconomic variables in commodity price forecasting by proposing a __micro-founded business cycle model incorporating endogenously determined commodity prices__.

Through applying a set of orthogonality conditions and sign constraints implied by their model, the authors decompose the sources of exogenous variation in their model into direct and indirect components. The **direct components** (DC) include shocks which directly shift supply and demand curves for commodities and thus would affect commodity prices, all else constant. By contrast, the **indirect components** (IC) affect commodity prices only through affecting aggregate output – and thus provoking an indirect pass-through into commodity market conditions. The __authors estimate that their identified indirect components factor accounts for around 2/3 of the variance in commodity prices__.

### Finding ‘proxies’ for macro conditions

Baumeister, Korobilis and Lee (2020) evaluate the success of a number of macroeconomic conditions proxies in predicting real Brent prices and refiner acquisition costs out-of-sample.

- Kilian (2009) prosed as a proxy for macro conditions
**changes in real shipping costs expressed in deviations from a linear time trend**. This proxy is meant to capture the cyclical component of demand for industrial commodities Although the predictor can lead to improvements over the random walk out of sample for refiner acquisition costs, it fails to do so for the real price of Brent. - An alternative factor can be constructed by
**extracting the principal components of a shipping cost dataset**. A factor-based approach has the advantage of filtering out commodity-specific noise from the proxy. For forecasts for refiner acquisition costs, this alternative shipping cost factor outperforms Kilian’s original factor. For forecasts for Brent, this alternative achieves substantial improvements over Kilian’s indicator, at all horizons. - To test the predictive power of
**world industrial production**, the authors use the index developed by Baumeister and Hamilton (2019). A VAR forecasting model incorporating this proxy considerably outperforms alternatives using either of the shipping-based indicators when forecasting both Brent prices and the refiner acquisition costs and leads to substantial improvements over the random walk at multiple horizons when forecasting the Brent price. - A real non-energy commodity price factor can be constructed from the
**first principal component of 23 industrial and agricultural commodity prices**. This factor is competitive with world industrial production at short horizons and somewhat better at longer horizons. - A global steel production factor can be constructed using the
**common component of monthly data on crude steel production**by country from the World Steel Organisation’s yearbook. At short horizons, this index underperforms industrial production and the commodity price factor when forecasting both the RAC and Brent, but outperforms Kilian’s index. At long forecasting horizons, this index is competitive with world industrial production.

### A new indicator for oil price pressures

“The question to which we now turn is whether global economic conditions as they relate to energy markets can be represented by any narrow set of variables or combinations thereof or whether there is value in diversifying the basket of variables to include new categories that cover additional dimensions of the global economy.” [Baumeister, Korobilis and Lee (2020)]

The authors propose __a new indicator for global economic conditions from a set of 16 variables relating to eight broad categorie__s, including:

**real economic activity**, incorporating data including world industrial production and the Conference Board Leading Economic Index,**commodity prices**, incorporating the real price of copper – which has been highlighted in a number of studies as a barometer for future global growth,**financial conditions**data, including on exchange rate fluctuations and equity returns,**expectations proxies**, including the Michigan consumer survey and data on the term structure of WTI futures,**uncertainty proxies**, including the geopolitical risk index developed by Caldara and Iacoviello (2018),**transportation data**, including vehicle registrations and traffic volume proxies,**weather data**, including information on temperatures and El Nino.**energy-related indicators**, including measures of production and distribution.

A global economic conditions indicator is defined as the first principal component of the dataset.

The global conditions indicator can be decomposed into its four main contributors, real economic activity, uncertainty measures, financial conditions and transportation indicators.

This factor-augmented model produces forecasts for real Brent prices which marginally underperform versus the other indicators evaluated above. However, the forecasts for global petroleum *consumption* dramatically improve over an AR(12) benchmark at all forecast horizons.

The authors’ Bayesian estimation strategy recovers an entire distribution of forecasts for each horizon, as opposed to a point estimate, which allows for the estimated conditional probability of oil prices rising outside a specific range to be calculated. Using this property of their estimation strategy, a measure for upside and downside oil price pressures is constructed by calculating the probability that oil prices, on average over the next year, fall above/below the range of prices seen over the last twelve months.

“In the aftermath of the Asian financial crisis, the probability that oil prices would fall below the lowest price over the past 12 months remained consistently high at 40% over a period of two years before plummeting to zero in early 2000. After oil prices reached a record low in December 1998, the likelihood of upward price pressures spiked resulting in a 55% probability that the Brent price will on average surpass its highest value during the past year over a 12-month horizon. At the onset of the Great Recession the average probability that the Brent price would exceed the previous year’s price maximum over the next 12 months dropped from around 40% to essentially zero, while chances of prices falling below the lower threshold jumped up to 50% followed by an all-time high of staying within the new lower price range of close to 80% in mid-2009. From 2012 onward, the odds that the Brent price would drop below its most recent lower bound gradually increased reaching a peak of 70% in early 2015. Starting in 2016 it becomes more and more likely that oil prices will top the price ceiling in place during the preceding 12 months in the coming year. As of August 2019, the price pressure measures indicate on average a 20% probability of the Brent price exceeding the recent upper threshold of $81 and a 30% probability of falling below the recent lower threshold of $57 in the period up to August 2020.”

### References

Alquist, R., Bhattarai, S. and Coibion, O., 2020. Commodity-price comovement and global economic activity. *Journal of Monetary Economics*, *112*, pp.41-56.

Baumeister, Christiane, and James D. Hamilton (2019). “Structural Interpretation of Vector Autoregressions with Incomplete IdentiÖcation: Revisiting the Role of Oil Supply and Demand Shocks,” American Economic Review 109(5): 1873-1910

Baumeister, Christiane and Lutz Kilian, 2012. “Real-Time Forecasts of the Real Price of Oil,” Journal of Business and Economic Statistics, 30(2), 326-336.

Baumeister, C., Korobilis, D. and Lee, T.K., 2020. *Energy Markets and Global Economic Conditions* (No. w27001). National Bureau of Economic Research.

Caldara, Dario, and Matteo Iacoviello (2018). “Measuring Geopolitical Risk,” mimeo, Federal Reserve Board of Governors.

Ferrari, D., Ravazzolo, F. and Vespignani, J., 2019. Forecasting energy commodity prices: a large global dataset sparse approach.