DData - Asian Development Bank
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
dData
Our nowcasting challenge definition
Estimate all macroeconomic variables, for all countries, daily, using all available information
• We aim to estimate all four tracks of the challenge (production, consumption, trade and prices).
By cross-cutting tracks we build holistic economy estimates using relations between indicators
• We assess all countries in Asia and the Pacific, including those with extremely poor data as
these are often the most vulnerable, and where nowcasting is most useful
• We nowcast variables daily in order to integrate available data at its most granular level
• We create a scalable platform so that additional information and new sources will continually
improve our estimates
The idea To predict macroeconomic indicators by linking hundreds of diverse datasets and combining them to accurately nowcast economic measurements Our technology links and aggregates economic datasets from a range of primary sources, from the World Bank’s Development Indicators to daily commodity price streams, helping users achieve faster, timely and more accurate insights into the world’s most pressing economic challenges Our pilot test nowcast GDP, consumption, exports and inflation in the Philippines, Pakistan and Myanmar during the 2008 - 2009 financial crisis. We built a scalable solution, identified a community interested engaging with it, and are looking to implement our solution in late 2020 and early 2021.
The methodology
• We convert each dataset into a common format that is comparable between sources. Where we have
1
multiple sources of the same data point, we select the highest quality estimate
• We then use the available historic data to look at the relationships between variables. Not all data is
2
available for each day and we predict missing variables daily, backfilling the missing data points
• We then identify what information we have at the current date and use the historic relationship of those
3
variables with the variable of interest to nowcast the current economic measurement and predict future
trends
1 2 3Vignette for nowcasting the financial crisis We use data from the 2008 - 2009 financial crisis to assess the effectiveness of our predictions We assume knowledge of the sporadic data prior to 01/01/2008 and predict across the crisis for all the countries in the ADB dataset from that date, adding daily data as it would be available from that date We test to see if our model can accurately nowcast the macroeconomic shocks for the next 12 -months and then predict the recovery until 01/01/2012
Link data sources at different frequencies
To predict the crisis, we use ~150 macroeconomic variables at daily, monthly and annual frequencies
Daily stock exchange API Annual macroeconomic databases
• NASDAQ Global and Asia Indices • World Bank Development Indicators
• Commodity markets (Agriculture & metals) • WB Worldwide Governance Indicators
• WB Enterprise Surveys and Doing
Business
• WEF Global Competitiveness Index
Monthly indexes API • IMF World Economic Outlook
• S&P index • Economist Intelligence Unit
• Financial rates including interest, earnings • Harvard Atlas of Economic Complexity
• Trade (Freight and Price) • Penn World Tables
• Economist Big Mac Index • Quality of Government DatasetHigh data countries – The Philippines
The Philippines is an
example of a county with
Philippines high data availability
01/01/2008
3500 10
9 - GDP and consumption
3000
8
data are available
2500
quarterly
7
- Inflation and export data
6
2000 are available monthly
5
1500 Our backfill model finds a
4
3
high degree of variation and
1000
tracks indicators effectively
2
500 creating daily trend
1 estimates from the low-
0 0 frequency point data
1/1/00 1/1/01 1/1/02 1/1/03 1/1/04 1/1/05 1/1/06 1/1/07 1/1/08 1/1/09 1/1/10 1/1/11 1/1/12
Quarterly real GDP Quarterly Consumption Expenditure Monthly Merchandise Exports We predict at a quarterly
Monthly Inflation Predicted GDP Predicted Consumption time-frame using 24
Predicted Exports Predicted Inflation variables, including shipping
indices amongst other
macroeconomic indicators.The model accurately nowcasts near-term output
Monthly Real GDP Monthly consumption expenditure
1800 1800
Quarterly Consumption in Bn Pesos
Quarterly Output in Bn Pesos
1600 1600
1400 1400
1200 1200
We find that the model accurately
1000 1000 nowcasts the dramatic fall in exports
800 800
600 600 during 2008 - 2009 as well as the first
400
200
400
200
year of real GDP and consumption
0 0 expenditure
10/10/06 2/22/08 7/6/09 11/18/10 4/1/12 10/10/06 2/22/08 7/6/09 11/18/10 4/1/12
Prediction ADB Dataset Prediction ADB Dataset
The model predicts some of the cost-
push inflation spike but fails to assess
Monthly merchandise exports Monthly Inflation its magnitude, rising only from 4% to
4000 14 5% rather than the peak of 12%
Quarterly Exports in Mn USD
Quarterly Inflation in Pct
3500 12
3000 10
2500
8
Overall the model nowcasts effectively
2000
1500
6 and has some success in prediction
1000 4
except for inflation magnitudes
500 2
0 0
10/10/06 2/22/08 7/6/09 11/18/10 4/1/12 10/10/06 2/22/08 7/6/09 11/18/10 4/1/12
Prediction ADB Dataset Prediction ADB DatasetMedium data countries - Pakistan
Pakistan Pakistan is an example of a
10000000
01/01/2008
12
county with medium data
availability
9000000
10
8000000
- GDP and consumption
7000000
8
data are available yearly
6000000 - Inflation and export data
5000000 6
are available monthly
4000000
4
Our backfill model finds a
3000000 lower degree of variation
2000000
2
and but still tracks indicators
1000000 effectively creating a smooth
0 0
trend out of infrequent GDP
1/1/00 1/1/01 1/1/02 1/1/03 1/1/04 1/1/05 1/1/06 1/1/07 1/1/08 1/1/09 1/1/10 1/1/11 1/1/12 and consumption data
Yearly real GDP Yearly Consumption Expenditure Monthly Merchandise Exports
Monthly Inflation Predicted GDP Predicted Consumption
In Pakistan only 14 variables
Predicted Exports Predicted Inflation
are available on a quarterly
basis to make predictionsThe model nowcasts well but fails to track variation
Yearly Real GDP Yearly consumption expenditure
10000000 9000000
Quarterly Output in Mn Rupees
Quarterly Consumption in Mn Rupees
9000000 8000000
8000000 7000000 We find that the model has a low degree
7000000
6000000
6000000
5000000
of variation in its nowcasting and
5000000
4000000
4000000 prediction failing to rise and fall at the
3000000
3000000
2000000 2000000
required magnitudes
1000000 1000000
0 0
10/10/06 2/22/08 7/6/09 11/18/10 4/1/12 10/10/06 2/22/08 7/6/09 11/18/10 4/1/12 The nowcasting of the first 6 – 12
Prediction ADB Dataset Prediction ADB Dataset months appear accurate however, the
model fails to predict the future
Monthly merchandise exports Monthly Inflation acceleration in Pakistan growth, or
3000000 30
exports and has difficulty with inflation
magnitudes
Monthly Exports in Th USD
Monthly Inflation in Pct
2500000 25
2000000 20
1500000 15 Overall the model is useful for
1000000 10 nowcasting in medium data countries
500000 5 but not appropriate for prediction
0 0
10/10/06 2/22/08 7/6/09 11/18/10 4/1/12 10/10/06 2/22/08 7/6/09 11/18/10 4/1/12
Prediction ADB Dataset Prediction ADB DatasetLow data countries - Myanmar
Myanmar is an example of a
county with low data
Myanmar availability
01/01/2008
18000
16000
- GDP and consumption
data are available yearly
14000 - Export data are available
12000 monthly
- Inflation data is not
10000
available before 2008
8000
Our backfill model finds a
6000
high degrees of variation
4000 and performs well at tracking
the scarce dataset including
2000
within year variation outside
0 of the initial dataset
1/1/00 1/1/01 1/1/02 1/1/03 1/1/04 1/1/05 1/1/06 1/1/07 1/1/08 1/1/09 1/1/10 1/1/11 1/1/12
Yearly real GDP Yearly Consumption Expenditure Monthly Merchandise Exports In Myanmar only 8 variables
Predicted GDP Predicted Consumption Predicted Exports are available on a quarterly
basis to make predictionsIdentifies within year variation not in original dataset
Yearly Real GDP Yearly consumption expenditure
45000 30000
Yearly Consumption in Bn MMK
Yearly Output in Bn MMK
40000
35000
25000 The model’s prediction identifies within
30000
25000
20000 year variation but fails to predict
20000
15000
Myanmar’s doubling of reported GDP
15000 10000
10000
between 2009-10, instead predicting an
5000
5000 economic slump
0 0
10/10/06 2/22/08 7/6/09 11/18/10 4/1/12 10/10/06 2/22/08 7/6/09 11/18/10 4/1/12
Prediction ADB Dataset Prediction ADB Dataset The nowcast of the first 6-months
appear accurate and useful as data is
Monthly merchandise exports
unavailable, however, predictions past
1200
this point appear very different to the
Monthly Exports in Mn USD
1000
trend
800
600 Overall the prediction accuracy appears
400 much weaker but potentially helpful in
200 nowcasting the first 6-months of missing
0
10/10/06 2/22/08 7/6/09 11/18/10 4/1/12
macroeconomic data
Prediction ADB DatasetAssessment and near-term implementation plan Our initial estimates are very rough and rely heavily on the high -frequency financial datasets, with more time and more computational capacity we would have linked more daily data and achieved a greater predictive accuracy We aim to improve on this initial vignette when modelling the COVID-19 crisis by: 1. Extending the model to this year, increasing the training dataset size from 8 to 20 years 2. Quality check existing values with alternative and subnational datasets 3. Automating the data formatting and nowcasting pipeline at all frequencies 4. Increase the number and heterogeneity of the data sources available We plan to incentivize people to add data to the predictive model by creating a marketplace to buy and sell data in the common linked format. We are aiming to go-live with the dData exchange during late 2020 and early 2021
The marketplace – https://atddata.com
We have created a peer-to-peer data sharing platform that crowdsources alternative datasets
1. The platform stitches together datasets and provides a single searchable output
2. Anyone can publish data and monetise its use through setting a price in Ocean tokens. We
encourage a free-open, collaborative effort with a Wikipedia-like delivery model
3. Transactions quick and secure through online payment based on a e-token wallet system
2 3
1You can also read
