Stacking: private API
PortfolioOptimisers.BaseStackingOptimisationEstimator — Type
abstract type BaseStackingOptimisationEstimator <: NonFiniteAllocationOptimisationEstimatorAbstract supertype of the optimisers that combine several inner optimisers through one outer optimiser.
A stacking optimiser treats the portfolio of each inner optimiser as one synthetic asset, and an outer optimiser allocates across these synthetic assets. So the returns of the inner portfolios set their weights, and no fixed rule averages them. This is the stacked generalisation of [246], with portfolios in place of predictors.
No method of the library dispatches on this type, so a subtype implements no method for it. Stacking is its one subtype.
Related
References
- [246] D. H. Wolpert. Stacked generalization. Neural Networks 5, 241–259 (1992).
PortfolioOptimisers.stacking_td_defaults — Function
stacking_td_defaults(
) -> @NamedTuple{pe::EmpiricalPrior{PortfolioOptimisersCovariance{Covariance{SimpleExpectedReturns{Nothing, Nothing, Nothing}, GeneralCovariance{SimpleCovariance, Nothing, Nothing}, FullMoment, Nothing, Nothing, Nothing}, MatrixProcessing{Posdef{UnionAll, @NamedTuple{}}, Nothing, Nothing, Nothing, NTuple{4, Symbol}}, Nothing}, SimpleExpectedReturns{Nothing, Nothing, Nothing}, Nothing, Nothing, Nothing, Nothing, Nothing}, opti::NoDefault, opto::NoDefault, wf::IterativeWeightFinaliser{Int64}}
Return the static defaults of the Stacking fields that can hold a TimeDependent schedule.
The constructor reads them when it tests each entry of a schedule, and time_dependent_field_defaults returns them. opti and opto are required and have no static default, so their entry is NoDefault. A schedule in one of them must carry its own default, or it cannot run outside a fold loop. The entries of pe and wf are their keyword defaults. The tuple holds no entry for wb, fees, sets, scale and fb, because their static default is nothing.
Returns
NamedTuple: The defaults ofpe,opti,optoandwf.
Related
PortfolioOptimisers.erase_optimiser_vector — Function
erase_optimiser_vector(opti::AbstractVector) -> Any
Erase the element type of a vector of inner optimisers to OptE_Opt_TD, so that the type of a Stacking does not change with the types of its inner optimisers.
A literal such as [MeanRisk(), TimeDependent(…)] infers the element type AbstractEstimator. The VecOptE_Opt_TD bound refuses that type, but it accepts each element. So the keyword constructor of Stacking calls this method on opti. The positional constructor calls it too, so a vector that a lens writes, or that a schedule resolves to, leaves with the same element type.
The element type is erased and not narrowed. A narrow element type, the union of the types of the elements, puts the type of each inner optimiser into the type of the Stacking, and optimise then compiles again for each combination of inner optimisers. The loop over opti dispatches on each element at run time, and that dispatch costs nothing next to the inner solve.
Algorithm
- Check that every element of
optiis anOptE_Opt_TD. - Convert
optito aVector{OptE_Opt_TD}. A vector of that type returns unchanged.
The method for a TimeDependent returns the schedule unchanged.
Arguments
opti: Vector of inner optimisers, or aTimeDependentschedule for the whole field.
Validation
- Every element of
optiis anOptE_Opt_TD. The method throws anArgumentErrorotherwise.
Returns
- The
Vector{OptE_Opt_TD}of the same elements, or the schedule itself.
Related
PortfolioOptimisers.is_time_dependent — Method
is_time_dependent(opt::Stacking) -> Any
Return whether the Stacking holds a TimeDependent schedule that a fold loop must resolve.
Returns true when time_dependent_fields reports a field of opt, or when opt.opti, opt.opto or opt.fb holds a schedule at any depth.
Related
PortfolioOptimisers.reset_time_dependent_estimator — Method
reset_time_dependent_estimator(opt::Stacking) -> Stacking
Replace each TimeDependent field of a Stacking with its static default.
The method does not recurse into the optimisers that opt holds. The inner cross-validation of opt resolves the element schedules of opti for each fold, and each full-sample inner solve resets its own schedules in its own _optimise. An element schedule of opti stays in place, because opti is then a vector and not a schedule. A field schedule of opti binds :outermost, and the method replaces it with its default.
Arguments
opt: The stacking optimiser.
Returns
Stacking: The optimiser with no schedule in its own fields.
Related
PortfolioOptimisers.assert_special_nco_requirements — Method
assert_special_nco_requirements(opt)Checks that an optimiser meets the requirements of an inner optimiser of NestedClustered.
The fallback method checks nothing. An estimator with a requirement adds a method, for example Stacking.
Arguments
opt: An optimisation estimator or result, a schedule of them, or a vector of them.
Returns
nothing.
Related
PortfolioOptimisers.assert_special_nco_requirements_stacking_opti — Function
assert_special_nco_requirements_stacking_opti(
opti::AbstractVector
)
Refuse a precomputed result among the inner optimisers of a Stacking that a NestedClustered holds.
NestedClustered solves its inner optimiser on each cluster, and its outer optimiser on the synthetic assets of the clusters. So a Stacking inside it runs on a universe that it does not know in advance. A NonFiniteAllocationOptimisationResult holds weights over the universe of its own solve, so it cannot give the returns of a synthetic asset on a new universe. A Stacking that runs alone accepts a result in opti.
assert_special_nco_requirements calls this method on opti, and on each entry and the default of a schedule in opti.
Arguments
opti: Inner optimisers of theStacking.
Validation
- No element of
optiis aNonFiniteAllocationOptimisationResult. The method throws anArgumentErrorotherwise.
Returns
nothing.
Related
References
- [246]
- D. H. Wolpert. Stacked generalization. Neural Networks 5, 241–259 (1992).