Upload
others
View
20
Download
1
Embed Size (px)
Citation preview
1©Hilltop Economics, LLC
Please do not redistribute without permission
The ESF Model(ESF: Econometric Semiconductor Forecast)
Real Time Forecasting Lessons-Learned
AUBER ANNUAL MEETINGOctober 2018
Duncan MeldrumHilltop Economics LLC
*Based on Consensus Economics, Inc. CONSENSUS FORECASTS®
2©Hilltop Economics, LLC
Please do not redistribute without permission
Outline
• Forecast process
• Lessons Learned
• Semiconductor MSI measure– Data analysis – pch, unit roots, trends, etc.
– Seasonal adjustment
• Model philosophy
• Model system & key final demand drivers
• Alternative Models
• ESF historical performance
3©Hilltop Economics, LLC
Please do not redistribute without permission
Forecast Process
• Analyze data after new quarterly MSI data release• Evaluate past forecasts against latest actual data• Validate model and equation structures with latest
actual data• Update Consensus Forecasts for World Economy, key
U.S. inputs• Update non-Consensus forecasts• Produce preliminary forecast for team discussion• Create alternative model quarterly forecast (ARIMA,
Book-to-bill, MIDAS models)• Develop final forecast for MSI
Process
4©Hilltop Economics, LLC
Please do not redistribute without permission
Lessons Learned• I will (almost) always be “wrong” – (but not by much).
– Tetlock: Superforecasting: The Art and Science of Prediction is right: open-mindedness is critical.
• Judgment and flexibility regarding modeling techniques and forecast preparation are essential
• The process is as important as the modeling– A model I can understand almost always produces better forecasts than a
model I can’t understand
• Tracking the forecast and adjusting for errors is key to a good forecast– One quarter ahead: use time-series & Midas to adjust– Downplay anecdotal evidence not substantiated by any data, but… – Pay attention to comments back from the users of the forecasts
• For the MSI and semiconductor industry measures:– Seasonal adjustment gets the variability– Final demand is more important than immediate markets
• There is too much “noise” in the immediate markets• Data availability in immediate markets is questionable
5©Hilltop Economics, LLC
Please do not redistribute without permission 5
400
800
1,200
1,600
2,000
2,400
2,800
3,200
3,600
96 98 00 02 04 06 08 10 12 14 16 18
SEMIMSI
Fitted Trend (Log)
Million Square Inches
Semiconductor Measure: MSI From SEMI.ORGSemiconductor wafers shipped for semiconductor applications
95Q1-18Q2: 2.2% Average quarter change
12.5% Standard deviation (+/-)5.8% Trendline growth rate per year
6©Hilltop Economics, LLC
Please do not redistribute without permission 6
Distribution of MSI Quarterly Percent Changes
0
4
8
12
16
20
24
28
-40 -30 -20 -10 0 10 20 30 40 50 60 70 80
Series: PCH_SEMIMSI
Sample 1995Q1 2018Q2
Observations 94
Mean 2.271178
Median 2.636729
Maximum 79.36170
Minimum -36.33527
Std. Dev. 12.39935
Skewness 1.947261
Kurtosis 18.68398
Jarque-Bera 1022.855
Probabilit
7©Hilltop Economics, LLC
Please do not redistribute without permission 7
Census X-13 Analytical Results:MSI: 95-18Q2
8©Hilltop Economics, LLC
Please do not redistribute without permission 8
Seasonal Factor Analysis
500
1,000
1,500
2,000
2,500
3,000
3,500
96 98 00 02 04 06 08 10 12 14 16 18
SEMI MSI
Seasonally Adjusted (X-13)
Mil
lio
n S
qu
are
In
ch
es
Source: Semi.org
Hilltop Economics LLC seasonal adjustment
Model as of Aug 2018
9©Hilltop Economics, LLC
Please do not redistribute without permission 9
-10
-8
-6
-4
-2
0
2
4
6
Q1 Q2 Q3 Q4
Means by Season
SEMI MSI SEASONAL IMPACT
History 1995Q1 to 2018Q2
Pe
rce
nt
Imp
ac
t o
f S
ea
so
na
lity
Becoming Less Seasonal
Model as of Aug 2018
10©Hilltop Economics, LLC
Please do not redistribute without permission
Model Philosophy• Semiconductors a dynamic but maturing industry
– Technological change (“supply side”) no longer creates large step-changes in MSI demand.
– Chips ubiquitous in products– Products ubiquitous in global regions
• Final Demand a valid driver of MSI– Demand driven by “final demand” – consumers and businesses using
technology products • Consumption • Investment
– Consumers & businesses buy technology products based on economic factors, so MSI can be modeled similar to other dynamic but maturing materials and products.
• Final Demand is best forecast by the global Consensus Forecasts of key (85) economies– “Wall Street credibility”
11©Hilltop Economics, LLC
Please do not redistribute without permission 11
Some Key Immediate End-use Markets(The “typical” industry forecast drivers)
0
100
200
300
400
500
08 09 10 11 12 13 14 15 16 17 18 19
Smart Phones
Smart Phones (Scenario 2)
Source:
History: Estimated from Gartner & IDC news releases
Forecast: Hilltop Economics, September 2018 Update
Mil
lio
n U
nit
s
Estimated Global Smartphone Sales
Annual Percent Change
2016 2017 2018 2019 2020
5.2% 2.6% - 0.8% -1.7% 4.5%
40
50
60
70
80
90
100
04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19
Estimated PC Units
September 2018 Update
Source: Hilltop Economics
Shaded area indicates main period of tablet penetration
Mil
lio
n U
nit
s
Estimated Global PC Sales
Annual Percent Change
2016 2017 2018 2019 2020
-7.0% -2.3% 1.1% -0.4% -2.3%
0.0
0.4
0.8
1.2
1.6
2.0
2.4
2.8
3.2
3.6
96 98 00 02 04 06 08 10 12 14 16 18
Servers
Source: Gartner news releases
Mil
lio
ns
12©Hilltop Economics, LLC
Please do not redistribute without permission 12
MSI VS “World” Real GDP.94 Correlation From 95Q1 to 18Q2, Level Log-Log
Note, however, penetration moderation 2011-2016
3,500
3,000
2,500
2,000
1,500
1,000
500
80,000
70,000
60,000
50,000
40,000
96 98 00 02 04 06 08 10 12 14 16 18
SEMIMSI
World Real GDP
BN
20
10
US
$M
illio
n S
qu
are
In
ch
es
94% Correlation
13©Hilltop Economics, LLC
Please do not redistribute without permission 13
MSI VS Real Investment.952 Correlation From 95Q1 to 17Q3, Level
Slightly better correlation, slightly lower growth multiple
18,000
16,000
14,000
12,000
10,000
8,000
3,500
3,000
2,500
2,000
1,500
1,000
96 98 00 02 04 06 08 10 12 14 16
SEMI MSI
Real Investment, 45 Consensus Countries
Million Square InchesBillion 2010 US$
95-17Q3
95.2% correlation
MSI Multiple: 1.7X
14©Hilltop Economics, LLC
Please do not redistribute without permission
MSI Block: The Core Model Structure
• Error Correction Model (ECM) system
– MSI seasonal adjustment model (Census X-13 Tramo/Seats) to capture “typical” seasonal patterns.
– Long run MSI-final demand embedded equation
– Short run MSI-dynamic cyclical equation to capture excursions from underlying long-run relationships
• Driven by exogenous inputs and models of final demand, plus any necessary supply-side developments.
Model
15©Hilltop Economics, LLC
Please do not redistribute without permission 15
Consensus Forecasts:
Global Real GDPInvestment Consumption
US Real GDP & Potential GDPReal Investment Real ConsumptionInflation & Interest Rates
ESF Model System(ESF: Econometric Semiconductor Forecast)
MSI Block ofEquations
PC & Smartphones
US Investment & ConsumptionEquations
US Tech Goods
Seasonal Factors(Census X-13)
Spending Propensity:Business TechConsumer Tech
Tablet Penetration, WinXP, Win10 Timing, Replacement
Cycles, Financial crash behavior
Economic PolicyUncertainty
(US, EU, China)
Inventory-Shipments:* PCs & Peripherals
* METI Wafers
Semiconductor:Capacity: Total, 300
Reclaim: % Total
Src: Hilltop Economics
16©Hilltop Economics, LLC
Please do not redistribute without permission
Long Run Embedded Equation
• Seasonally-adjusted MSI modeled as a function of key inputs:– “Global” ex US investment and consumption– US technology goods spending (business & consumer)– Global PC and smartphone unit sales– Economic Policy Uncertainty (Global)– Interest rates– Semi wafer capacity estimates (total, VNAND
introduction)– METI I-S ratios (wafers)– Behavioral dummies: financial collapse & 9/11
Model
17©Hilltop Economics, LLC
Please do not redistribute without permission
Short Run Cycle Equation
• Quarterly rate of change in MSI modeled as a function of key inputs:– Deviation from long run embedded equation estimates– Rate of change in key long run inputs:
• Global final demands• US Tech• PCs & smartphones• Economic policy uncertainty
– Rate of change in US inventory-shipments ratio for computers & peripherals, METI I-S ratio
– Special events measures (9/11, Financial Crisis onset, Win10)
– Rate of change in SEMI capacity measure and impacts from 300mm and VNAND intorductions (lagged one Qtr)
Model
18©Hilltop Economics, LLC
Please do not redistribute without permission 18
Unit Root Test: Log(SEMI MSI) does not
have a unit root
19©Hilltop Economics, LLC
Please do not redistribute without permission 19
The Long-run Embedded MSI EquationAug 18: Adjusted R-square: .986 D-W: 1.76 SER: 4.2% MAPE: 3.9%
Model as of Aug 2018
20©Hilltop Economics, LLC
Please do not redistribute without permission 20
The Long-run Embedded MSI EquationNov 17: Adjusted R-square: .986 D-W: 1.76 SER: 4.2% MAPE: 3.9%
21©Hilltop Economics, LLC
Please do not redistribute without permission 21
Long-run EmbeddedEquation
Model as of Aug 2018
22©Hilltop Economics, LLC
Please do not redistribute without permission 22
Dynamic Error Correction Equation
Model as of Aug 2018
23©Hilltop Economics, LLC
Please do not redistribute without permission 23
Dynamic Fit 98Q3-18Q2
Model as of Aug 2018
24©Hilltop Economics, LLC
Please do not redistribute without permission 24
In-sample Forecast
Model as of Aug 2018
25©Hilltop Economics, LLC
Please do not redistribute without permission
Consensus Forecasts®
©Consensus Economics, Inc. produces four major reports covering the world economy:
1) Developed World
2) Asia-Pacific
3) Latin America
4) Eastern Europe
Hilltop Economics, LLC uses these reports with permission to create a global economic forecast of real GDP, real investment and real consumer spending.
ESF customers are encouraged to subscribe to Consensus Forecasts ® for the macroeconomic information needs.
Process
26©Hilltop Economics, LLC
Please do not redistribute without permission
Typical Country Output:
Key concepts for ESF forecast:
• Gross Domestic Product
• Personal Consumption
• Business Investment
• Consumer Prices
• Producer Prices
Interest Rates (on second page of document, not shown)
Process
27©Hilltop Economics, LLC
Please do not redistribute without permission 27
-1
0
1
2
3
4
5
2014 2015 2016 2017 2018 2019 2020
June Forecast
September Forecast
World* Real GDP Growth
2017: 3.1% 2018: 3.1% 2019: 3.0% 2020: 2.7%%
Ch
an
ge
*W orld: 85 major economies,
Forecast: W orld Bank, Hilltop Economics based on
Consensus Forecasts, June 18/September 18
Examine Changes to Consensus Forecasts
*World: 85 countries in Consensus Forecasts. Source: November Consensus, IMF, Hilltop Economics
Process
28©Hilltop Economics, LLC
Please do not redistribute without permission 28
Increased Uncertainty Raises Downside Risk
Process
0
50
100
150
200
250
300
350
400
98 00 02 04 06 08 10 12 14 16 18
U.S.
Europe
China
Source: Scott Baker, Nicholas Bloom and Steven J. Davis
at www.PolicyUncertainty.com. Data through Sep 18
Ind
ex
3M
MA
<- Financial Crisis
Economic Policy Uncertainty
29©Hilltop Economics, LLC
Please do not redistribute without permission 29
Conservative Projections of U.S. Behavior 15Q4
.40
.44
.48
.52
.56
.60
.64
6
7
8
9
10
11
12
96 98 00 02 04 06 08 10 12 14 16
Consumer (LEFT SCALE)
Business Investment
Percent of Consumption/Investment
Spent on Technology Goods (US)
Process
30©Hilltop Economics, LLC
Please do not redistribute without permission 30
Conservative Projections of U.S. BehaviorA judgmental input based on recent trends
Process
7
8
9
10
11
12
.2
.3
.4
.5
.6
.7
90 92 94 96 98 00 02 04 06 08 10 12 14 16 18 20
Business Investment
Consumer Spending
Based on U.S. Bureau of Economic Analysis data to 18Q2
Shaded areas indicate U.S. recessions (NBER)
Perc
en
t In
vestm
en
tP
erc
en
t Co
nsu
mp
tion
31©Hilltop Economics, LLC
Please do not redistribute without permission 31
Updated Monthly For U.S. NIA Releases
Process
-20
-15
-10
-5
0
5
10
15
06 07 08 09 10 11 12 13 14 15 16 17 18 19
U.S. Business Investment (P&E)
U.S. Technology Equipment
%C
H V
S Y
R A
GO
, 4
QM
A
History: U.S. Bureau of Economic Analysis
Sep 2018 Forecast: Hilltop Economics
Shaded areas indicates official U.S. recession
32©Hilltop Economics, LLC
Please do not redistribute without permission
Tracking the Alternatives
Actual
“ESF”
Ranges
ARIMA
2 Std Dev
Book-to-Bill: Latest 3 months
MIDAS: Latest 3 months
Add factors
Seasonal factors
Previous month’s forecast
33©Hilltop Economics, LLC
Please do not redistribute without permission
Tracking the Alternatives
Actual
“ESF”
Ranges
ARIMA
2 Std Dev
Book-to-Bill: Latest 3 months
MIDAS: Latest 3 months
Add factors
Seasonal factors
Previous month’s forecast
34©Hilltop Economics, LLC
Please do not redistribute without permission
Tracking the Alternatives
Actual
“ESF”
Ranges
ARIMA
2 Std Dev
Book-to-Bill: Latest 3 months
MIDAS: Latest 3 months
Add factors
Seasonal factors
Previous month’s forecast
35©Hilltop Economics, LLC
Please do not redistribute without permission
Tracking the Alternatives
Actual
“ESF”
Ranges
ARIMA
2 Std Dev
Book-to-Bill: Latest 3 months
MIDAS: Latest 3 months
Add factors
Seasonal factors
Previous month’s forecast
36©Hilltop Economics, LLC
Please do not redistribute without permission
Tracking the Alternatives
Actual
“ESF”
Ranges
ARIMA
2 Std Dev
Book-to-Bill: Latest 3 months
MIDAS: Latest 3 months
Add factors
Seasonal factors
Previous month’s forecast
37©Hilltop Economics, LLC
Please do not redistribute without permission 37
Graphic Examination of Forecasts
2,400
2,600
2,800
3,000
3,200
3,400
3,600
3,800
4,000
I II III IV I II III IV I II III IV I II III IV I II III IV
2016 2017 2018 2019 2020
SEMIMSI
SEMI MSI (Scenario 2) - ECM
SEMIMSI_2M+SEMIMSI_2S
SEMIMSI_2M-SEMIMSI_2S
SEMIMSI_2M+SEMIMSI_2S*2
SEMIMSI_2M-SEMIMSI_2S*2
SEMIMSI_F_18Q3 - ARIMA
38©Hilltop Economics, LLC
Please do not redistribute without permission 38
How Well Does this Forecast Process Work?The August ESF forecast a weaker Q3 than developed
2,400
2,600
2,800
3,000
3,200
3,400
II III IV I II III IV I II III IV I II III IV I
2015 2016 2017 2018 2019
SEMI MSI
ESF 2018 Q2
Mil
lio
n S
qu
are
In
ch
es
ESF 18Q2 over-forecast by 1.5%
(3207 forecast vs 3160 actual)
History: SEMI to 18Q2
Forecast: ESF 2018 Q2 (May)
Performance
39©Hilltop Economics, LLC
Please do not redistribute without permission 39
How Well Does this Forecast Process Work?Last November’s Forecast VS Actual
2,000
2,200
2,400
2,600
2,800
3,000
IV I II III IV I II III IV I II III IV I II III IV
2012 2013 2014 2015 2016
SEMI MSI to 16Q3
2015 Q4 Forecast
Performance
40©Hilltop Economics, LLC
Please do not redistribute without permission 40
Now 23 Forecasts Since the First Forecast (of 2012Q4)
2,000
2,200
2,400
2,600
2,800
3,000
3,200
3,400
II III IV I II III IV I II III IV I II III IV I II III IV I II III IV I II III IV I
2012 2013 2014 2015 2016 2017 2018 2019
SEMIMSI Q4 2012 ESF Forecast MSI
Q1 2013 ESF Forecast MSI Q2 2013 ESF Forecast MSI
Q3 2013 ESF Forecast MSI Q4 2013 ESF Forecast MSI
SEMI MSI Feb 14 ESF SEMIMSI_MAY14
SEMIMSI_AUG14 SEMIMSI_NOV14
SEMIMSI_FEB15 SEMIMSI_MAY15
SEMIMSI_AUG15 SEMIMSI_NOV15
SEMIMSI_FEB16 SEMIMSI_MAY16
SEMIMSI_AUG16 SEMIMSI_NOV16
SEMIMSI_FEB17 SEMIMSI_MAY17
SEMIMSI_AUG17 SEMIMSI_NOV17
SEMIMSI_FEB18 SEMIMSI_MAY18
Mil
lio
n S
qu
are
In
ch
es
Performance
41©Hilltop Economics, LLC
Please do not redistribute without permission 41
Summary Performance
Cumulative Error on LEVELS (FCST VS ACTUAL)
1 Quarter Ahead
2 Quarters Ahead
3 Quarters Ahead
4 Quarters Ahead
8 Quarters Ahead
ESF: Average % Error
-1.1% -0.8% -0.9% -0.8% -0.6%
ESF: Mean Absolute % Error
2.4% 2.4% 2.6% 2.6% 2.6%
ARIMA Average % Error
-0.5% -0.9% -1.2% -1.3% -2.4%
# Forecasts 23 22 21 20 16
Performance
42©Hilltop Economics, LLC
Please do not redistribute without permission
Lessons Learned• I will (almost) always be “wrong” – (but not by much).
– Tetlock: Superforecasting: The Art and Science of Prediction is right: open-mindedness is critical.
• Judgment and flexibility regarding modeling techniques and forecast preparation are essential
• The process is as important as the modeling– A model I can understand almost always produces better forecasts than a
model I can’t understand
• Tracking the forecast and adjusting for errors is key to a good forecast– One quarter ahead: use time-series & Midas to adjust– Downplay anecdotal evidence not substantiated by any data, but… – Pay attention to comments back from the users of the forecasts
• For the MSI and semiconductor industry measures:– Seasonal adjustment gets the variability– Final demand is more important than immediate markets
• There is too much “noise” in the immediate markets• Data availability in immediate markets is questionable