Tommelfingerregler vs. optimering
Transcription
Tommelfingerregler vs. optimering
Wealth Planning and Machine Learning Kourosh Marjani Rasmussen, Schantz ([email protected]) DTU ([email protected]) 1. september 2015 1 Normative vs. Positive Economics • Normative Economics: “1.000.000 people can’t be wrong.” • Positive Economics: “Yes, they can, and I can prove it.” 2 Life cycle financial planning Phase 1 Phase 2 25 Depletion Consolidation Build up Age: Phase 3 45 65 85 3 Planning ingredients The household economy Disposable income 4 Optimization Model (Positive Economics) Objective: Find the largest disposable income for the entire lifecycle given some percentage reduction at a given year 5 The objective functions Objective Function 1: Respect the disposible income up to the retirement year. Thereafter, maximize the smallest disposible income for a given year. 𝒎𝒂𝒙𝒊𝒎𝒊𝒛𝒆 𝐹𝑖𝑥𝑒𝑑𝐴𝑁𝑃 w.r.t. 𝐹𝑖𝑥𝑒𝑑𝐴𝑁𝑃 ≤ 𝐴𝑁𝑃𝑡 = 𝑋𝑡 𝐴𝑁𝑃𝑡 𝑎𝑙𝑝ℎ𝑎𝑡 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 𝑖𝑓 𝑅𝑒𝑡𝑖𝑟𝑒𝑚𝑒𝑛𝑡𝑇𝑖𝑚𝑒𝑡 = 1 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 𝑖𝑓 𝑊𝑜𝑟𝑘𝑇𝑖𝑚𝑒𝑡 = 1 6 The objective functions Objective Function 2: Obtain the maximum disposible income for the entire period. 𝒎𝒂𝒙𝒊𝒎𝒊𝒛𝒆 𝐹𝑖𝑥𝑒𝑑𝐴𝑁𝑃 w.r.t 𝐹𝑖𝑥𝑒𝑑𝐴𝑁𝑃 ≤ 𝐴𝑁𝑃𝑡 𝑎𝑙𝑝ℎ𝑎𝑡 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 7 The objective functions Objective Function 3: Maximize the disposible income prior to retirement while respecting a given target during the retirement period. 𝒎𝒂𝒙𝒊𝒎𝒊𝒛𝒆 𝑌 w.r.t 𝐴𝑁𝑃𝑡 = 𝑋𝑡 + 𝑌 𝐴𝑁𝑃𝑡 ≥ 𝑇𝑎𝑟𝑔𝑒𝑡𝑡 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 𝑖𝑓 𝑊𝑜𝑟𝑘𝑇𝑖𝑚𝑒𝑡 = 1 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 𝑖𝑓 𝑅𝑒𝑡𝑖𝑟𝑒𝑚𝑒𝑛𝑡𝑇𝑖𝑚𝑒𝑡 = 1 8 Social Benefits 1) 𝐼𝑛𝑐𝑜𝑚𝑒 = 𝑃𝐼1𝑡 + 𝑃𝐼2𝑡 + 𝑃𝐼3𝑡 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 2) 𝑃𝐼1𝑡 ≤ 𝐴𝑡 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 3) 𝑃𝐼2𝑡 ≤ 𝐵𝑡 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 4) 𝑃𝐼3𝑡 ≤ 𝑃𝐼3_Bin𝑡 ∙ 𝑀 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 5) 𝐵𝑡 ∙ (1 − 𝑃𝐼3Bin 𝑡 ) ≥ 𝐵𝑡 − 𝑃𝐼2𝑡 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 6) 𝑆𝑜𝑐𝑖𝑎𝑙𝐵𝑒𝑛𝑒𝑓𝑖𝑡𝑡 = 𝑀𝑎𝑥𝑃𝑎𝑦𝑚𝑒𝑛𝑡𝑡 − 𝑃𝐼2𝑡 ∙ 𝑅𝑒𝑑𝑢𝑐𝑡𝑖𝑜𝑛𝑅𝑎𝑡𝑒𝑡 ∀ 𝑡 ∈ 𝑡𝑖𝑚𝑒 9 Constraints • Pensions Different tax regimes for different pensionschemes Specific cashflow restrictions • Social benefits Six different types Interaction with different sources of income • Loans Mortgage loans Consumption loans Bank loans Reverse mortgages • Tax • Companies Income tax Capital gain tax Stock tax Pension tax Tax rebates 10 We can model all this! Model Statistics: The objective function is to maximize the family´s disposable income with respect to +300 generic constraints The model is a mixed integer program solved with GAMS/CPLEX. Approximate number of constraints: 100.000 Approximate number of variables: 200.000 Approximate number of binary variables: 10.000 11 Or can we? 12 And can we get people to use it? 13 But can we model human behavior We learn gradually – so can our models. Ex: Google face recognition network trained on millions of pictures. 14 Mini Project: Wealth Planning & Machine Learning Machine learning can cluster clients into similar groups. Human interaction with the system will be captured. (Normative Economics) The optimization model will be enhanced with time. (Hybrid Economics: Do what works and teach what is right!) 15 KOUROSH MARJANI RASMUSSEN M: +45 40866164 E: [email protected] SCHANTZ A/S KIGKURREN 10 DK-2300 KØBENHAVN S T: +45 3332 1984 WWW.SCHANTZ.COM 16