Presentation is loading. Please wait.

Presentation is loading. Please wait.

Ana Galvão, Anthony Garratt and James Mitchell. September 2017

Similar presentations


Presentation on theme: "Ana Galvão, Anthony Garratt and James Mitchell. September 2017"— Presentation transcript:

1 Ana Galvão, Anthony Garratt and James Mitchell. September 2017
Density Forecasting in Practice: UK output growth and inflation. The WBS Forecasting System Ana Galvão, Anthony Garratt and James Mitchell. September 2017

2 WBS Forecasting System - Galvao, Garratt, Mitchell
Outline (1) What do we do? (2) How do we do it? (3) How have we done? WBS Forecasting System - Galvao, Garratt, Mitchell

3 WBS Forecasting System: What do we do?
Point forecasts, the traditional focus of macroeconomic forecasters, are best interpreted as the central points of ranges of uncertainty. The Warwick Business School Forecasting System quantifies and communicates these forecast uncertainties by producing probabilistic forecasts. Addresses the issue of model uncertainty by using combinations of many models. Calendar-year forecasts of output Growth and inflation are released every quarter at the day of second UK GDP release. WBS Forecasting System - Galvao, Garratt, Mitchell

4 WBSFS August 2017 Forecasts
These are based on a combination of statistical forecasting models: They are judgement-free calendar-year forecasts! WBS Forecasting System - Galvao, Garratt, Mitchell

5 WBSFS August 2017 Forecasts
We provide probability forecasts for events: low growth and inflation outside the targeted interval. The main purpose is to aid decision making. Year Real GDP Growth (%, p.a.) CPI Inflation (%, p.a.) Prob(growth<0%) Prob(growth<1%) Prob(growth<2%) Prob(letter) Prob(CPI<1%) Prob(CPI>3%) 2017 <1% 2% 88% 10% 2018 5% 19% 45% 33% 23% WBS Forecasting System - Galvao, Garratt, Mitchell

6 WBSFS August 2017 Forecasts
Another objective of the WBS forecasting system is to provide benchmark and judgement-free probability forecasts. Forecasts for 2018: Real GDP Growth (%, p.a.) CPI Inflation (%, p.a.) Point Forecast Prob. of a higher outturn Bank of England 1.68 0.63 2.52 0.39 HMT Panel 1.40 0.71 2.60 0.36 WBS Forecasting System - Galvao, Garratt, Mitchell

7 WBS Forecasting System - Galvao, Garratt, Mitchell
How do we do it? Model Set Simple Statistical Autoregressive Model: AR: autoregressive model of order 2 estimated with quarterly growth rates; density forecasts computed by bootstrap (parameter and forecasting uncertainties). A Mixed Frequency Regression to exploit Monthly Info: MIDAS: direct regression of quarterly targets on monthly indicators, allowing for one month lead depending on publication delays; Beta weighting function; density forecasts computed by fixed-regressor bootstrap (both uncertainties). We consider 13 indicators so we have 13 single indicator MIDAS models. WBS Forecasting System - Galvao, Garratt, Mitchell

8 WBS Forecasting System - Galvao, Garratt, Mitchell
Model Set A “Macroeconomic” BVAR Model: BVAR_macro: Bayesian VAR model of real GDP and CPI quarterly macro series (Consumption, Investment, Hours, Real Wages, Bank Rate). Variables in levels, p=4; Normal/Wishart to obtain Density Forecasts. “Leading Indicator” BVAR Models: BVAR_medium: Bayesian VAR model of real GDP and CPI indicators (as in the MIDAS models). Normal/Wishart to obtain Density Forecasts. 8 different specifications: p=1,…,4 x Levels/Differences. A Time-Varying BVAR with Stochastic Volatility: TVP_VAR: A model of output growth, inflation and the short- rate. p=1. Gibbs + MH step draws to obtain density forecasts [H. Mumtaz code]. WBS Forecasting System - Galvao, Garratt, Mitchell

9 WBS Forecasting System - Galvao, Garratt, Mitchell
Our Information Set Name Publication Delay Output Growth 56 days (2nd) Inflation 14 days Consumption 56 days Investment Real Wages 75 days Hours Bank rate no delay Industrial Prod. 45 days Business Confidence Employment Unemployment Consumer Confidence House Prices Stock Prices Exchange Rates 10 days House Prices _2 Short rate 9 days Yield Spread Retail Prices Oil Prices 29 days WBS Forecasting System - Galvao, Garratt, Mitchell

10 WBS Forecasting System - Galvao, Garratt, Mitchell
Why combine? Point forecasts are more robust to instability/structural breaks since models’ relative forecasting performance may change over time. It might be easier to find density forecasts that approximate well the true data density. Equal weighting might be an inadequate combination method: it does not down weight bad forecasters … Weighting scheme based on past performance may perform better than equal weighting if forecasters have very different performance. Evaluation with quarterly fixed horizon forecast suggested that log pool performs better than the linear pool. WBS Forecasting System - Galvao, Garratt, Mitchell

11 WBS Forecasting System - Galvao, Garratt, Mitchell
Combination Method Total number of models/experts: 24. Logarithm opinion pool: 𝑑𝑒𝑛𝑠 𝑡+ℎ|𝑡 = 𝑖=1 𝑚 ( 𝑑𝑒𝑛𝑠 𝑡+ℎ|𝑡 𝑖 ) 𝑤(𝑖) , where rescaling is required. Selecting Model’s (Expert’s) weights: Proportional to the relative forecasting performance up to t measured by the logscore. This uses a rolling window of 12 quarterly observations. Weights estimated to fixed horizon quarterly forecasts and using the second monthly release of GDP as the actual value (real-time!) WBS Forecasting System - Galvao, Garratt, Mitchell

12 Model weights: time-varying, specific to period and variable forecast
2017q2 (h=3,7) 2017 –output 2018 -output inflation inflation ar 0.057 0.042 0.131 0.052 bvarmacro 0.054 0.047 0.067 0.051 tvpvar 0.020 0.025 0.028 0.024 ip 0.041 0.000 0.038 bus conf 0.044 0.001 emp 0.061 0.056 0.075 unem 0.017 0.045 cons conf 0.027 0.031 house pr 0.037 stock pr 0.046 exc rates 0.053 0.043 house pr 2 0.064 0.032 short rate 0.018 0.030 0.040 yield sp 0.039 0.036 retail prices oil prices L-bvarmed_p1 0.026 L-bvarmed_p2 L-bvarmed_p3 0.050 0.023 0.035 L-bvarmed_p4 0.014 D-bvarmed_p1 0.217 D-bvarmed_p2 0.162 D-bvarmed_p3 0.117 0.048 D-bvarmed_p4 0.109 2017q3 (h=2, 6) 2017 -output 2018 -output inflation inflation ar 0.0542 0.0437 0.7325 0.0593 bvarmacro 0.0427 0.0468 0.0011 0.0552 tvpvar 0.022 0.0264 0.0159 0.0167 ip 0.0576 0.0477 0.0362 bus conf 0.0461 0.0454 0.04 emp 0.0596 0.0567 0.0556 unem 0.0473 0.051 0.0448 cons conf 0.0307 0.0155 0.028 house pr 0.05 0.0425 0.0428 stock pr 0.053 0.0353 exc rates 0.0503 0.0475 0.0267 house pr 2 0.0621 0.056 0.0326 short rate 0.0348 0.0242 0.0424 yield sp 0.0458 0.0422 0.0393 retail prices 0.039 0.0455 0.0401 oil prices 0.0299 0.0132 0.0467 L-bvarmed_p1 0.0201 0.0511 0.1358 0.0397 L-bvarmed_p2 0.0313 0.0004 0.0333 L-bvarmed_p3 0.0303 0.0482 0.0311 L-bvarmed_p4 0.029 0.048 0.0271 D-bvarmed_p1 0.047 0.0391 0.0569 0.0555 D-bvarmed_p2 0.0412 0.0379 0.032 0.0572 D-bvarmed_p3 0.038 0.0559 D-bvarmed_p4 0.0381 0.0389 0.0243 0.0586 WBS Forecasting System - Galvao, Garratt, Mitchell

13 WBS Forecasting System - Galvao, Garratt, Mitchell
How have we done? Formal evaluation difficult because Just 12 forecasting rounds, three of which (for 2017) with respect unknown future outcomes Fixed event nature of forecast events – therefore effectively have just three/two observations per event More formal analysis undertaken when choosing models. Therefore here we tabulate and plot some probabilities ties and outcomes, as well as the history of our benchmarking exercise WBS Forecasting System - Galvao, Garratt, Mitchell

14 WBS Forecasting System - Galvao, Garratt, Mitchell
Table 1: Output Growth, Probabilities and Outturns Probability of current year growth Probability of next year growth <0% <1% <2% Obs. 2014q1 0.00 3.07 0.03 0.07 0.22 2.19 2015q1 0.01 0.05 0.24 0.06 0.14 0.29 1.81 2015q2 0.27 0.13 2015q3 0.02 0.08 0.25 2015q4 0.28 2016q1 0.10 0.45 2016q2 0.43 2016q3 2016q4 0.23 WBS Forecasting System - Galvao, Garratt, Mitchell

15 WBS Forecasting System - Galvao, Garratt, Mitchell
Table 2: Inflation, Probabilities and Outturns Probability of current year inflation Probability of next year inflation <0% <1% <2% Obs. 2014q1 0.00 1.00 1.41 0.12 0.41 0.76 0.04 2015q1 0.44 0.87 0.31 0.58 0.82 0.65 2015q2 0.84 0.99 0.50 2015q3 0.29 2015q4 0.19 0.74 0.96 2016q1 2016q2 2016q3 2016q4 WBS Forecasting System - Galvao, Garratt, Mitchell

16 WBS Forecasting System - Galvao, Garratt, Mitchell
Table 3: Probability of NOT being in inflation target range of 1-3% i.e. writing a letter Probability in current year Probability next year Prob. Obs. 2014q1 0.00 Hit 1.41 0.46 No hit 0.04 2015q1 0.87 0.63 0.65 2015q2 0.99 0.52 2015q3 1.00 2015q4 0.73 2016q1 2016q2 2016q3 2016q4 WBS Forecasting System - Galvao, Garratt, Mitchell

17 WBS Forecasting System - Galvao, Garratt, Mitchell

18 WBS Forecasting System - Galvao, Garratt, Mitchell

19 WBS Forecasting System - Galvao, Garratt, Mitchell

20 WBS Forecasting System - Galvao, Garratt, Mitchell

21 WBS Forecasting System Future
Provide quarterly releases of calendar-year probabilistic forecasts based on state-of-art forecasting models. Forecasts are published at: casting/ Update model set/combination weights based on comprehensive research on density forecasting combination. Keep forecasts judgment free so we can evaluate the extension of judgment in alternative published forecasts. WBS Forecasting System - Galvao, Garratt, Mitchell

22 APPENDIX: Evaluating WBSFS Forecasts
We use a long out-of-sample evaluation period. Forecasts are computed from 1999Q1 to 2012Q3. Real-time exercise: only data actually available at the time forecast is computed (ONS monthly real-time dataset of real GDP; and for other macro variables) and takes into account publication delays of series not subject to revisions. Quarterly forecast horizons from h=1 (current quarter forecast) up to h=8 (two-year ahead forecast). [2015 forecast computed in Feb 2015 requires forecasts for h=1 up to h=4; if computed in Nov 2015, only h=1 forecast is required since previous quarters are available]. WBS Forecasting System - Galvao, Garratt, Mitchell

23 Evaluating WBSFS Forecasts
Measure of point forecast accuracy: root mean squared forecast error (RMSFE). Measures of density forecasting accuracy: logscore and continuous ranking probability score (CRPS). The measures above allow us to test for statistical significant differences in forecasting performance. Measures of calibration of density forecasts. Being well- calibrated means approximating well the true data density. We use tests based on probability integral transforms (PITs). WBS Forecasting System - Galvao, Garratt, Mitchell

24 Comparing Accuracy with BofE
Output Growth ( N=55) Point Forecast Density Forecast – CRPS score Density Forecast – Log Score 1 2 4 8 BofE 0.526 0.80 1.588 2.364 0.307 0.439 0.822 1.17 4.493 4.704 5.823 6.474 AR 1.10 1.34 1.29 0.97 1.02 1.27 1.23 0.96 1.05 1.00 0.89 BVARmed 0.82 1.06 1.01 1.03 0.98 0.94 Comb; m=2; 1.07 1.12 0.93 1.11 Comb; m=5 1.19 1.32 1.18 1.04 0.99 Comb; m=24;log 0.88 1.24 1.21 1.15 1.14 1.09 1.08 Inflation(N=35; from 2004). Point Forecast Density Forecast – CRPS score Density Forecast – Log Score 1 2 4 8 BofE 0.184 0.483 1.088 1.512 0.130 0.283 0.625 0.845 3.55 4.214 5.127 5.449 AR 2.11 1.54 1.16 0.60 1.65 1.44 1.12 0.68 1.01 0.93 MIDAS 1.75 1.30 1.18 0.86 1.37 1.20 0.84 1.06 1.05 0.95 BVARmed 2.34 1.84 1.64 1.48 1.74 1.66 1.55 1.34 1.13 1.25 1.07 1.03 Comb; m=2 2.19 1.35 0.98 1.50 1.26 0.90 0.96 Comb; m=5 2.15 1.77 1.47 1.59 1.52 1.14 1.10 1.08 1.02 Comb; m=24;log 1.19 0.87 1.21 1.17 WBS Forecasting System - Galvao, Garratt, Mitchell

25 Calibration of Density Forecasts
Output Growth ( N=55) Mean PITs: 0.5 (Berkowitz Test) Std Deviation PITs: (Knuppel Test) 1 2 4 8 BofE 0.41 0.40 0.35 0.34 0.29 0.27 0.26 0.24 AR 0.45 0.37 0.21 BVARmed 0.49 0.51 0.28 Comb_2 0.47 0.48 0.46 0.44 Comb_5 0.43 0.25 0.22 Comb_24_log 0.39 Inflation(N=35; from 2004). Mean PITs: 0.5 (Berkowitz Test) Std Deviation PITs: (Knuppel Test) 1 2 4 8 BofE 0.34 0.37 0.45 0.49 0.30 0.36 0.41 AR 0.48 0.44 0.29 0.28 0.26 0.19 MIDAS 0.52 0.51 0.56 0.23 BVARmed 0.50 0.24 Comb_2 0.27 0.21 Comb_5 0.55 0.57 0.60 0.61 0.22 Comb_24_log 0.53 WBS Forecasting System - Galvao, Garratt, Mitchell

26 Main Evaluation Messages
The BofE Inflation Report Forecasts of Output Growth and Inflation perform well if compared with statistical models in particularly when predicting current year inflation. Their predictive density however are not good approximations of the true data density. By combining a set of statistical models, we can obtain well- calibrated densities for current year forecasts of output growth and inflation. [we can also to get well-calibrated two-years ahead inflation forecast]. Regular releases of well-calibrated predictive densities of output growth and inflation, which help decision making, is one of the main purposes of the WBS Forecasting System. WBS Forecasting System - Galvao, Garratt, Mitchell


Download ppt "Ana Galvão, Anthony Garratt and James Mitchell. September 2017"

Similar presentations


Ads by Google