Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Automated Supply-Use Balancing in the United Kingdom: A New Approach Introduction • SU balancing is the reconciliation of data sources to provide a GDP estimate • It should depend on the quality of each data source • Traditionally, this has been done manually • But better computers are helping NSIs automate this • So how to specify the algorithm? Quadratic Programming - A Model • QP is a well-established algorithm with multiple easily-available software implementations • But what constraints and objective to use? • Usual model – Weight the objective (squared adjustments) for relative quality – Accounting constraints – Optional soft constraints for economic ratios Literature Review • Chen (2012) presents this model in matrix algebra terms • Key points: Covariance matrix between initial estimates and true values must be known; initial estimates must be unbiased estimators of true values; if so, model produces unique best solution • Similar conclusions in Bikker et al (2013) and van Tongeren and Picavet (2016) The UK Position • We do not know the covariance matrix (Chen) • Nor do we know the standard errors of the underlying dist. of the initial estimates – or even whether they are normally distributed (Bikker et al, van Tongeren and Picavet) • So can we develop a model suited to these conditions? An Alternative Model • Consider the matrices and vectors underlying the SU system • For each cell, we should be able to allocate upper and lower bounds (eg, the UK agriculture industry produces less than £100 trillion for each product, and more than £0 for the agriculture product) • Real bounds are much narrower • So we can apply these as additional constraints in a LP/QP model • No weights on the objective Comparison of Models • This model does not guarantee a unique best solution (indeed, the solution may not be unique – but this is unlikely) • And, if the constraints are not set correctly, the problem can be infeasible • These are consequences of the unavailable information • So we can’t guarantee the outcome will be as good as the other model – but we don’t have to make its assumptions Testing • Testing was done on past UK SU tables • Upper and lower bounds based on manual adjustments – if a cell had been manually adjusted by + or – x%, the upper and lower bounds were set as +/- x% • Some different treatment based on different manual procedures (eg trade in goods) Results • See tables in paper • For the 4 main SU aggregates (GDP(P), GDP(I), total supply, and total demand) total adjustments were lower for all except GDP(I) • At the lower level, the lower manual GDP(I) adjustments were the result of higher (in absolute terms) adjustments netting off • So results are positive Conclusions • Automation of the SU process is viable • ONS is still not able to deliver the more ambitious aims of the alternative model • That requires more work to establish the needed metadata • This work should be a top priority for ONS and the other NSIs working on automated balancing