- 1. Price Sensitivity (0.1) Towards 1:1 pricing Modern/dynamic pricing approaches applied to personal unsecured loans (PUL). Paul Golding, Dec 2018 @pgolding
- 2. Preface ā¢ We want to predict a borrowerās sensitivity to price and use this information to optimize prices to achieve maximum expected profit (or some other business goal - TBD) ā¢ This deck is a āthink aloudā conversation to gain a common understanding of the principles and possible solutions to the pricing sensitivity problem. It also poses questions that need answers to move forward with building a pricing model. ā¢ The deck contains a mostly mathematical tour of the principles because the goal is to discover a computational method(s) that might achieve 1:1 pricing segmentation via a fully automated system. ā¢ We heavily lean upon the work of R. Phillipsā (Pricing Credit Products) and referenced materials. A literature review suggests that his work is a definitive baseline. In places, we have formulated the math in a more accessible fashion.
- 3. Credit Risk 1. Risk is a variable cost 2. Risk is correlated with price sensitivity: riskier customers tend to be less price sensitive 3. Price influences risk: all else equal, raising prices leads to higher losses - price-dependent risk.
- 4. Terms ā¢ Exposure at Default (EAD) - remaining balance at default ā¢ Loss Given Default (LGD) = EAD/Total Loan Amount ā¢ Probability of Default (PD) = P(Default) | t<=loan_term ā¢ EAD and LGD are random variables ā¢ (Note: we largely stick to Phillipsā terminology to be consistent with his theoretical foundations. Variation from more standard terminology is noted.)
- 5. Expected Loss We care about expected values of random vars PD = pd (t) t=1 T ā EEAD = pd (t) Ć EADt t=1 T ā PD ELGD = pd (t) Ć EADt t=1 T ā Ć LGDt pd (t) Ć EADt t=1 T ā EL = pd (t) Ć EADt t=1 T ā Ć LGDt ā“ EL = PD Ć ELGD Ć EEAD Probability of default at t=1 or t=2 or t=3 ā¦ or t=T Expected loss is major determinant of expected profitability of a loan and thereby influences price.
- 6. Simplified Pricing Threshold E Ļ ( )= 1ā PD ( ) p ā pc ( )ā PD Ć (1+ pc ) One-period supply capital rate pc One-period price p Expected profit for $1 1-period loan (to simplify understanding risk/price relationship): profit loss p > PD 1+ pc ( ) 1ā PD + pc cost of capital Risk-adjusted premium This only suggests a lower bound. It does not suggest the price: ā¢ Maximize profitability? ā¢ Maximize bookings subject to profit constraint? ā¢ Price sensitivity of customer Extremely simplified: ignores dynamics of payment/ default in multi-period loan and that PD is not independent of price at which loan oļ¬ered. But suļ¬cient to show the role of risk. A function of PD pc =0.1 Risk cut-oļ¬ equation:
- 7. Risk Prediction s x ( )= ln Pr{G | x} Pr{B | x} ā ā ā ā ā ā score of observables vector log odds (i.e. monotonic) modeled using ML (e.g. XGBoost) Presumably decision-tree based for reasons of interpretability (actual adverse actions methodology: see D. Paulsen). (Look ahead: can we get better sensitivity predictions if we donāt care about interpretability?) PĢ{B | x}
- 8. Risk Factors PĢ{B | x} ā¢ Risk prediction based upon borrower observables ā¢ But also the loan itself ā¢ Amount ā¢ Leverage (e.g. LTV, DTI) ā¢ Period (longer loans riskier due to information asymmetry, random shock etc.) ā¢ And price-dependent risk Not part of x Borrower dependent risk Product dependent risk
- 9. Price-dependent Risk If for all values of E Ļ ( )= 1ā PD ( ) p ā pc ( )ā PD Ć (1+ pc ) ā E Ļ ( )= 1ā PD p ( ) ( ) p ā pc ( )ā PD p ( )Ć (1+ pc ) P ā² D p ( )> 0 i.e. positive slope, or risk increases with price PD p ( )> p ā pc p +1 p ā„ pc then there is no price at which a loan is profitable
- 10. Price-dependent Risk ā¢ Fraud - fraudulent borrowers are insensitive to price ā¢ Adverse private information - e.g. unobservable knowledge of a borrowerās forthcoming redundancy ā¢ Competitive alternatives - less riskier customers have more options and so loans at a certain (higher) price will attract riskier borrowers ā¢ Behavioral factors - behavioral economics, mental accounting, time discounting etc. (Dynamic contextual factors unrelated to loan characteristics or borrower credit-worthiness attributes.) ā¢ Aļ¬ordability - e.g. DTI (capacity to pay)
- 11. Incremental Profitability ā¢ Optimal price imposes two questions: ā¢ How will profitability of a loan change with price? ā¢ How will demand for a loan change with price? ā¢ Profitability for loans non-trivial calculation due to: ā¢ Time value of money (discounted value of future costs/revenues) ā¢ Uncertainty: default, prepayment, utilization (for credit line) ā¢ We need a model for incremental profitability for PUL ā¢ What are the variable costs? (Holds, channels?) ā¢ What is the profit goal? (For whom?) ā¢ If we sell/securitize loans, how does this aļ¬ect optimization? ā¢ Longer loans are more profitable (via interest) ā¢ Large-portfolio holders should act more risk-neutral (assuming risk of loans are statistically independent across the portfolio) - see Phillips ā¢ Also: should pricing model take LTV into account - e.g. cross-selling other products? āØ (Segmentation question?) Action: construct the incremental profitability model
- 12. Price Response Curve ā¢ We want to know the price- response function per product per segment (vs. market- demand curve per product) - i.e. their Willingness To Pay (WTP) ā¢ Price optimization uses this along with expected incremental profit calculation to price for maximum expected total profitability (for a loan portfolio) d p ( )= DF p ( ) total applications (i.e. āmarket sizeā) take-up rate (i.e. WTP) price-response (i.e. ādemandā) Adapted from curve in this paper: Journal of Business Research Loan APR Loan APR F p ( )= f w ( )dw p ā ā« probability distribution of WTP WTP is probability of a customer paying at most p - i.e. note that it is the complementary cum. dist. func. or 1 - F(p) (often seen in the literature) D = d(0) - i.e. willing to buy at all (source: DTU ME) Additional source: BPO lecture notes. Nominal demand curve (closed form continuous) In general, d(p) = D ( 1-F(p) ) = D(F(infinity)-F(p)).
- 13. Some derivations d p ( )= DF p ( )ā ā² d p ( ) = āD ā² F p ( )= Df p ( ) F p ( )= f w ( )dw p ā ā« ā ā ā² F p ( )= f p ( ) For a given WTP curve, the corresponding density Slope -ve, hence introduce minus sign In some cases, we are interested in absolute slope of the take-up rate, in which case: ā² F p ( )ā ā² d p ( ) / D = f p ( ) Note: the derivative can be interpreted as the percentage willing to pay exactly p: ā² F p ( )ā F(p + h)ā F(p) F(p) For small h
- 14. Measuring Sensitivity Typical curve more like this Note: the curve is actually time-dependent, but this is not reflected in the baseline math. There are many diļ¬erent ways in which product demand might change in response to changing prices but all price-response functions assume: ā¢ non-negative (pā„0) ā¢ continuous (no gaps in market response to prices) ā¢ diļ¬erentiable (smooth and with well-defined slope at every point) ā¢ downward sloping (raising prices decreases demand) ā¢ notwithstanding ābehavioral economics eļ¬ectsā Additional source: DTU Management Engineering Most common measures are: ā¢ Slope - dā(p1) is the derivative of the price-response function at p1 ā¢ Hazard rate (slope/demand) ā¢ Elasticity: %demand change from 1% change in price: Īµ p1 ( )= ph p1 ( )= ā² d (p1)p1 / d p1 ( )= pf p ( )/ F p ( ) h p1 ( )= ā² d (p1) d p1 ( ) = f p ( )/ F p ( ) point elasticity We want to estimate this curve from any actual take-up data. āFailure rateā (See Yao) has important underlying properties when calculating optimal pricing.
- 15. Sensitivity and risk ā¢ As said, higher prices = higher default rates AND higher risk customers are less price sensitive (e.g. price elasticity higher for higher FICO) ā¢ But the two phenomenon are the same: if price sensitivity decreases with risk in a population, then the populate will demonstrate price-dependent risk. db p ( )= DbFb p ( ) D = Dg + Db dg p ( )= DgFg p ( ) #applicants take-up rate #goods #bads DR = Db / Dg + Db However, the rate is price dependent, so: DR(p) = DbFb p ( ) DbFb p ( )+ DgFg p ( ) Diļ¬erentiating wrt price, we get: D ā² R p ( )= (hg p ( )ā hb p ( ))DR p ( )(1ā DR p ( )) failure rate diļ¬erence between goods and bads hg p ( )= ā ā² Fg p ( )/ Fg p ( )= fg p ( )/ Fg p ( ) hb p ( )= ā ā² Fb p ( )/ Fb p ( )= fb p ( )/ Fb p ( ) via substitution, chain-rule and product rule
- 16. Sensitivity and risk D ā² R p ( )= z p ( )DR p ( )(1ā DR p ( )) z p ( )= hg p ( )ā hb p ( ) Diļ¬erence in hazard rate Price-dependent risk rate (or PDR rate) - i.e. rate at which risk changes with price commercial lending region for constant z(p) and z(p)>0 In this part of the curve, price-dependent risk more salient for higher-risk populations. āAs found in practice - i.e. eļ¬ect of price on risk much stronger in sub- prime than prime. Very little impact in super prime.ā What does it imply for pricing? Net interest income for a loan of value 1 (for simplicity) and loss-given default: Ī p ( )= (p ā c)dg p ( )ā ldb p ( ) 0 < l ā¤1 There is a positive number k such that z(p)>k for all p >=c in which case the lender cannot achieve a profit at any price. Whether or not it is profitable to oļ¬er a loan depends not only on bads:goods but on their diļ¬erence in price sensitivity as reflected in the diļ¬erential hazard rate z(p)
- 17. Pricing/Underwriting Ī p ( )= (p ā c)dg p ( )ā ldb p ( ) There is a positive number k such that z(p)>k for all p >=c in which case the lender cannot achieve a profit at any price. Whether or not it is profitable to oļ¬er a loan depends not only on bads:goods but on their diļ¬erence in price sensitivity as reflected in the diļ¬erential hazard rate z(p) k = hb p ( )ā hg p ( )
- 18. Estimating d(p) ā¢ We assume a data-driven approach based upon random testing (of prices), but we need to evaluate the current random-testing method/data ā¢ Are we storing all endogenous variables? (if any) ā¢ Do we have suļ¬cient test data for each segment? ā¢ (Note: we can model using available data and also estimate with piecewise numerical estimator - see later.) ā¢ Given the historical data, we now estimate a price-response model using ML (GLMs or ANN?) ā¢ Do we have endogeneity? ā¢ What might cause it? e.g. agent intervention in the sale of the loan
- 19. Model ā¢ Considerations ā¢ Feature selection (for price sensitivity) ā¢ Feature engineering overhead ā¢ Monotonicity - The assumption of pricing optimization is that prices change monotonically ā¢ Can we short-cut feature engineering stage with use of a novel DL method, like Deep Lattice Networks? ā¢ Pros: shortcut feature engineering, Tensorflow models, organizational learning ā¢ Cons: lack of proven field experience (with DLN). Expertise better used elsewhere (portfolio optimization?) ā¢ Updating: frequency and method (to track temporal changes) ā¢ To what extent can this be automated? ā¢ Should we include other product co-joint analysis? āØ (āPrice to enticeā as gateway to higher LTV.)
- 20. Pricing Segmentation M Product dimensions (e.g. amount, term) N Customer dimensions (e.g. scores, employment) observable/explainable Price - Price - - Price Price - - Price Price - - - Price NxM buckets (upper bound) logical limit 1:1 pricing or.. NxM dimensional surface whose geometry determines revenue/profit and should be correlated to: ā¢ incremental profitability ā¢ price sensitivity ā¢ Smaller (more targeted) buckets => profitability, but also complexity. What is the trade-oļ¬? (We need to do work to analyze the frontier of ROI trade-oļ¬ curves.) ā¢ All else being equal, charge lower prices to segments with: ā¢ higher inc. profitability ā¢ higher price sensitivity ā¢ other business incentive e.g. promote certain channels? (Again, what are the business constraints.) ā¢ Customer dimensions: ā¢ Risk, [age], loyalty, [geography] etc. ā¢ Product dimensions: ā¢ Loan amount, term ā¢ Term - should we consider a higher resolution of term (than just 3/5 year) to obtain better results? (And give a more satisfactory UX.)? Are there other ways to diļ¬erentiate product besides price? e.g. āno-charge term rescheduling or repayment rewardsā i.e. customer segmentation achieved via product segmentation correlated to price sensitivity
- 21. WTP: Pricing Demand Price Demand Price Demand Price p* p1* p2* c c c Single price p* missed revenue = A,B Two prices c<p1*<p2* missed = A,B,C 1:1 pricing missed = ā0ā ā¢ We now see the importance of predicting WTP = revenue/profit ā¢ However, WTP extends beyond risk-based customer variables e.g. to include things like financial sophistication, brand aļ¬nity, urgency, competitive environment etc. ā¢ Note: are some of these observable, say via random experiments in design (or other means). Do we capture all the necessary data? What other data sources can we use? ā¢ i.e. pricing (sensitivity) is not a ābackendā function only ā¢ Can we gather more user data? (e.g. for DM clients can we aļ¬ord āsecond stage?ā)
- 22. online tasks Dynamic segmentation Query $ Price Dims Optimization Prices for all segments offline tasks Optimization $ Oļ¬er Oļ¬er Application Application Analysts reviewers ā¢ d(p) can be continually computed (and āevent drivenā) ā¢ New data resolution (higher freq) ā¢ Market shifts (lower freq) ā¢ What are the operational implications of moving to a 1:1 model? ā¢ To what extent can the entire process be automated? ā¢ Optimization model (inc. freq) ā¢ Operational model (as an operationally acceptable rule- based system - what are the constraints?) ā¢ Can/do we capture all the data we need (e.g. UX state)? ā¢ What is the limit at which other approaches should be considered (due to behavioral economics factors)? ā¢ The exact nature and implementation of the current oļ¬ine tasks need to be audited and automation ROI calculated Compute power is cheap. If we can find an end-to-end computational solution, even highly complex (AI), it potentially pays! What are the strategic (tech) imperatives in a commodity PUL market?
- 23. Optimizing Prices (Simple loan) E Ļ p ( ) ( )= F p ( )[ 1ā PD ( )Ć p ā c ( )ā PD Ć l] Consider a single two-period loan (just makes math simpler initially) [And assume LGD independent of price, which it is not.] l = LGD + c [simplistic incremental profit] price response = F p ( )[ 1ā PD ( )Ć p ā c ā l o ā ā ā ā ā ā] o = (1ā PD) / PD odds that loan will not default loss given default cost of funds ā o* = LGD + c p ā c underwriting criterion for a price-taking lender - i.e. take any loans with odds > o* But we are interested in price-setting lender. In which case we want to charge a price p* that would maximize Here we omit the analytical solution for as there is no closed-form solution for any non-linear price response function. Rather, we can see that if we can compute then we can calculate the maximum using numerical methods. E Ļ p ( ) ( ) ā² E Ļ p ( ) ( ) F p ( )
- 24. Price-dependent risk Analytical solution for h p* ( )= 1 p* ā c ā l o ā² E Ļ p ( ) ( )= 0 But we have so far assumed that PD and loss are independent of price of loan, which is incorrect. If we now assume that we have a population Dg of good and Db bad customers. E Ļ p ( ) ( )= Fg p ( )Dg p ā c ( )ā Fb p ( )Dbl ā² E Ļ p ( ) ( ) ā² E =0 āÆ ā āÆāÆ hg p* ( )= 1 p* ā c ā fb p* ( ) fg p* ( ) ā ā ā ā ā ā l o0 ā ā ā ā ā ā This holds for price-dependent risk (diļ¬erent dist. for bads/goods) ā² E Ļ p ( ) ( ) ā² E =0 āÆ ā āÆāÆ hg pĢ ( )= 1 pĢ ā c ā l o pĢ ( ) ā ā ā ā ā ā If we assume the odds will not change (i.e. ignore price-dependent risk) BUT: we do expect goods have higher failure rate => fb p ( )/ fg p ( )> Fb p ( )/ Fg p ( ) And therefore (plugging into above equations): Which means the (amateur) lender who does not consider influence of price upon risk will set a price that is higher than the expert lender, which means higher losses and lower profit. Note: this aļ¬ects competitors profitability - see Nomis white paper. p* < pĢ population odds
- 25. Adaptive Optimization 1. Iterate k pricing adjustments Ī±1 > Ī±2 > Ī±3...> 0 2. Use 2-point price estimation (testing) to estimate hĢ pk ( )= 2 d1D2 ā d2D1 ( )/[Ī“ d1D2 + d2D1 ( ) ( )] 3. Calculate price gap g(pk ) = hĢ pk ( )ā1/(pk ā c ā l / o) 4. if ā Īµ ā¤ g(pk ) ā¤ Īµ set p* = pk and stop 5. if g(pk ) < āĪµ set pk+1 = pk +Ī±k and go to step 2 6. if g(pk ) > Īµ set pk+1 = pk āĪ±k and go to step 2 1/ p ā c ā l / o ( ) h p ( ) assumed unknown c + l / o r* Īµ Īµ p Raise Hold Lower inverse margin What if we donāt know h(p)?
- 26. Optimization Over Segments p* = argp max DF p ( )Ļ p ( ) ( ) ! p* = argp max DiFi pi ( )Ļi pi ( ) i=1 N ā ā£ " p = argp max DiFi pi ( )Ri pi ( ) i=1 N ā ! p = argp max DiFi pi ( )Ļi pi ( ) i=1 N ā + DiFi pi ( )Ri pi ( ) i=1 N ā ā” ā£ ā¢ ā¤ ā¦ ā„ Or with constraints: e.g Bounds, structural constraints, monotonicity of pricing, segment bounds, rate endings ! p = argp max Ī¦C ( DiFi pi ( )Īøi pi ( ) i=1 N ā ) ā” ā£ ā¢ ā¤ ā¦ ā„ c=1 M ā ā” ā£ ā¢ ā¤ ā¦ ā„ Subject to feasibility whilst trying to avoid post-hoc pricing adjustments (āoverlaysā) due to āshortcomingsā in the model. And subject to idealized segmentation (that avoids mixed [separable] failure curves with implied lack of global maximum). And subject to location of the portfolio on the eļ¬cient frontier (next). Where/what are the automation opportunities for constraint review and management? (assume independently solvable)
- 27. Efficient Frontier Return Risk ā¢ We can estimate the frontier and where our portfolio lies within it ā¢ But we may have chosen unrealistic (or too many) constraints ā¢ We might have chosen constraints outside of the model (i.e. tuned prices manually) ā¢ Optimal portfolio selection is an area of R&D opportunity (reinforcement learning, genetic algorithms, etc.) ā¢ Optimal segmentation is also an area for R&D opportunity Can I have the highest revenue with greatest profit and lowest risk please? No.