Lars Weighted Revenue.pptxVIP

Setting Weights on Objective Function for Long Channel ClientEfficient Frontier Analytic TeamJuly , 2010Question:What are the weights? Why we need set weights?How we set weights? What are Weights? Weighted Revenue = ω1 × Metric1 + ω2 × Metric2LeadCompleted ApplicationProportion between Metrics 100010100:1(0.25,0.75) 2507.5250:7.5 = ?(1,300)100030001000:3000=0.25:0.75It’s the proportion between the metrics that matters, not the weightsWhy we need to set the Weights? KeywordCostLeadFollow UpRealized RevenueROIloan requirements$1022,250220%Mortgages$814,120170%Mortgage rates$3802680 0% Lending mortgage $6010 0 0% The last two keywords should not be treated as revenue 0They should be treated differentlyTwo Ways to GoClient’s Way: Evaluate the value of every metric. Rely highly on intuition Include every metricOur Way: Catch the relationship between each metric in the channel to lift the performance of tail.How we lift tail performance by setting weights?Weights that best predict revenueKeywordLeadFollow UpRevenueEstimateRealized RevenuePrediction ErrorWeight85615---loanrequirements25114$22,368$22,9893%Home loans22812$23,053$21,544-7%Mortgage rates268$7,130 -Lending mortgage 10$85- We get estimations of the revenue of two tail terms. We are able to treat them differentlyHow do we come up with the best weights?Should Every Metric be Included ? Every metric contains value of other metrics, co-linear problem Every metric has some volatility. So more metrics, more noise.LARS Algorithm: Less is MoreStep 0: Revenue = 0 × Metric1+0 × Metric2 +0 × Metric3+ 0 × Metric4Predictive Error = 1 Step 1: Revenue = 0 × Metric1+0 × Metric2 +ω3 × Metric3+ 0 × Metric4Predictive Error =0.85 Step 2: Revenue = ω1 × Metric1+0 × Metric2 +ω3 × Metric3+ 0 × Metric4Predictive Error =0.75 Step 3: Revenue = ω1 × Metric1+ ω2 × Metric2 +ω3 × Metric3+ 0 × Metric4Predictive Error =0.78 Step 4: Revenue = ω1 × Metric1+ ω2 × Metric2 +ω3 × Metric3+ ω4 × Metric4Predictive Error =0.81