QM 323 Midterm

market basket analysis

association rules: make predictions based on correlations between purchases -a form of one to one marketing

one to one target marketing

attempts to predict which individual customers are most likely to purchase

binding constraint

bottleneck -- prevents you from further increasing your objective ex: diamond tire case

k center

centroids/means

attributes

characteristics that define aspects of a product or service ex: "brand origin" , "body style", "engine type", "price", "number of seats"

choice based conjoint vs. full profile conjoint

choice: represents more real world marketplace choices full profile: conjoint that we're used to

if two rules have equal lift

choose the one with greater antecedent

item set

collection of items selected from set of items in the store

euclidean distance

distance between two observations on CONTINUOUS variables -group observations into clusters with smallest possible euclidean distance -before calculating distance, calculate z-score

limits of k means clustering

doesn't tell which cluster to target/most valuable only form with measurable variables only tells which observations are similar to eachother

constraints in linear programming formulations

equations that must always hold true -- typically inequalities; only term on the right hand side is the parameter(a number) and all choice variables are on left hand side

cluster analysis

figuring out who will buy the product by grouping customers into as FEW market segments/clusters as possible (less costly this way) uses quantitative info on customer characteristics/observations makes large datasets more manageable can be used to identify outliers & to better UX

heuristic decision

flawed decision marking technique that seems logical but isnt

cost benefit analysis

general rule: target customer if the expected benefits of targeting > expected costs expected benefit = # of purchases * profit margin per unit expected cost = # of mailings sent out * cost per mailing breakeven= when cost and benefit are equal

decision pitfalls

heuristic decision, confirmation trap, overconfidence, overfitting

value of association rules based on:

how actionable it is -How well it explains the relationship between item sets -an association rule is useful if it is well supported and explain an important previously unknown relationship

profiles

hypothetical products or services described as bundles of attributes ex: an american car with with sedan body and gas engine

baseline option

in conjoint, it's the combo of attributes that you use as base for regression omit one attribute from each category, you get baseline set baseline to zero (not literally zero, idiot)

optimization model: intercept in demand

increase in intercept = outward shift in demand this will lead to increase in price

sensitivity analysis

measure goodness of fit with decile-wise lift chart or a cumulative gains chart lift by decile is calculated by sorting data from highest to lowest probability and counting purchasers in top decile cumulative gains for model

lift ratio

measure to evaluate efficiency of a rule; a measure of relative predictive confidence confidence / (support of consequent/total number of transactions) lift will stay same no matter which is antecedent and which is consequent if lift >1, then there is predictive power lift = 1, then there's no relationship if lift < 1 then products are negatively associated -with a lift of 2, "you are twice as confident or 1-time more likely to buy buns when purchasing hot dogs than someone who just buys buns randomly -If lift = .88, the probability that a person who buys a Red Sox shirt buys a Yankees shirt is only 88% of the probability that a random person buys a Yankee shirt. You are 12% less likely to buy a Yankees shirt given you bought a Red Sox shirt

K means cluster

method of creating cluster based on the distance from the k center( aka centroids/mean) with K as number of variables reliable when you know how many clusters you want and large data set

predictive analytics

models based on existing data to make predictions about what is likely to happen if a particular decision is taken ex: data mining, simulation to stress test a financial model), regression

support of item set

number of times item set occurs in database

parameters

numbers in equation such as objective equation parameters don't change when you change the choice variables

influence chart symbols

objective: hexagon (ex: gross profit) variable: circle (ex: revenue, VC, FC) decision: rectangle (price, demand) fixed input: triangle (size of target market, labor costs per unit, material costs per unit)

multiple regression analysis

of purchases and customer characteristics to predict which customers are most likely -analysis of purchases and customer characteristics to predict which customers are most likely to purchase a product.

fixed costs in optimization model

only a VERTICAL shift of the curve....optimal price does not depend on level of fixed costs

cumulative gains chart for model

part of sensitivity analysis display proportion of purchasers we can expect to gain from targeting a specific percent of customers using the model shows difference between no model analysis and using analysis ex: results of using a model to target customers show that by targeting the top 10% most likely buyers, we would gain almost 30% of total buyers. By targeting the top 30% most likely buyers, we would gain almost 60% of customers. LARGER the area under the curve(fatter the banana), the better the model is a predicting purchases

how to make sure association rules are useful

perform cluster analysis before market basket analysis

logit model

predicts the ODDS of an event odds ratio: probability event occurs divided by probability it doesn't occur =p/(1-p)

optimization model: slope of demand

product with demand that is less sensitive to changes in price has, all else equal, a higher optimal markup cost

part worth plot

put all options customer has on one graph so you can see better which they prefer and by how much

optimization models

quantitative models to help identify best course of action given a set of alternatives ex: given relationship between price and sales, what should price be? 1. unconstrained optimization 2. constrained optimization

Conjoint analysis

randomly put attributes into a grouping and responses would deciding what combo/types of attributes --ASKS what attributes will create the best bundle for my consumers? -different from to have or not to have products and services are "bundles" of attributes -NORMALIZE data first -consumers choose the product/service that delivers the greatest total value -we can decompose the product/service into the value of each subpart

importance of attribute aka "how important is this attribute in conjoint analysis?"

range of an attribute/(SUM of ranges across all attributes) yes, sum!! of all of them think of airplane snack example

range of attribute

range of attribute = max partworth - min partworth

likert scale

rates profiles in conjoint analysis based on ordinal scale E.g. "Please rate each of the following options by selecting the number on the 7-point preference scale (1 = lowest, 7 = highest) that best reflects the strength of your preference"

influence chart

simple diagram of a model shows: what outcome variables the model will generate and how outputs are calculated from inputs NOT an analytical technique just helps organize/plan model and bring clarity to model design process objective: hexagon variable: circle decision: rectangle fixed input: triangle

descriptive analytics

summarize data to give decision makers sense of data they're working with EX: word clouds, scatter plots, pivot tables, clustering

sensitivity analysis

target cells's

minimum support criterion

threshold below which we ignore item sets (ex: we may omit support less than 1%)

unconstrained optimization vs. constrained optimization

unconstrained: find max/min of a function -"what price should we charge to earn MOST profit" constrained: max/min something while satisfying a set of conditions -"how much should i invest to maximize returns while not incurring too much risk?"

predictive regression modeling

use multiple variables to predict probability of purchases and identify best customers to target

levels

values of each attribute ex: brand origin would be "american" "japanese"

overfitting

when a statistical model describes random noise rather than the true underlying relationship we are trying to capture in order to make predictions

confirmation trap

when we pay more attention to findings that confirm our prior beliefs and ignore the facts

issues with lift ratio

while lift ratio may look great, look at confidence and support of BOTH happening

z score

(given observations mean - mean of variable)/SD

caveat

-Association rule with a high lift ratio and low support -may still be useful if the consequent represents a very valuable opportunity

confidence

-conditional probability -helps identify reliable association rules -confidence will DEFINITELY change when order of consequent and antecedent changes support of {antecedent and consequent} /support of antecedent confidence of "if hot dogs, then buns" is 3/5 = 60%...this means that in 60% of transactions, someone who buys hot dogs is also buying buns

linear programming formulation

LP is used to find optimal values of choice variables when optimization is constrained -in LP problems, boths OBJECTIVE FUNCTION and CONSTRAINTS are linear -objective function and constraints are written as mathematical formulas containing choice variables and parameters -difference between a linear program and other optimization problems is the linearity of the objective constraints

BE Response rate

Responses*$6 - $.60*Mailings = 0 Responses/Mailings (BE Response Rate) = .60/6 = 10% This implies send mailings to anyone in the dataset who has an estimated probability of purchase greater than 0.10

associate rules

-uses info on products that are commonly bought together -if then statements -used for: product placement in stores; cross category and co-marketing promotions; product bundling; recommendation systems

steps in K means clustering

1. allocate observations at random k clusters 2. calculate z score for each characteristics of the observations 3. calculate cluster centroids (average values of the variables/the z scores) 4. Calculate euclidean distance of each observation to each cluster centroid 5. move each observation into cluster with closest centroid 6. Repeat 3-5 until no observations need to be moved

estimating purchase probabilities

1. assign similar customers to a segment by demographics or clustering & target that segment 2. use "one to one marketing" with market basket analysis or predictive regression modeling

analyzing purchase probabilities using regression

1. define objective: maximize profitability of campaign 2. develop predictive model: check residuals for each purchase probability 3. sort customers from highest probability to lowest probabilty(sort residuals) 4. perform cost-benefit anaylsis 5. perform sensitivity analysis

standardized optimization model

1. define single objective: "make most profits" 2. define choice variables: price 3. express objective as single equation thats a function of the choice variables & parameters profits = rev-TC= rev-(FC+VC) 4. express equations "all other things that must hold true and might affect your choice" aka constraints (demand curve, max working hours) 5. make best choice 6. evaluate solution. 7. perform sensitivity analysis

steps of conjoint analysis

1. identify set of revelant product/service attributes: don't use too many attributes (below 6); focus on the attributes which managerial decisions need to be made; no ambiguous description 2. select levels (value) for attributes: nothing highly unrealistic; no confounded levels; compare apples to apples 3. Create product profile: hypothetical products/services described as bundles of attributes 4. Collect data: using traditional paper and pencil questionnaires or online survey 5. estimate part worths (coefficient): estimate consumers' preferences based using linear regressions to link consumer's ratings to preference for various attribute levels - literally just plot out all options and where they fall on graph 6. Derive insight and make predictions -MAKE SURE baseline is set to 0 -partworths of other levels = deviations from this baseline profile

LP: advantages and disadvantages

Advantage: Problems with large numbers of variables and constraints are possible Disadvantage: Linearity is sometimes unrealistic (remember, though, that models are approximations)

conditional probability

CONFIDENCE P (B|A) - confidence = P(AB) - support of A & B / P(A) - Support of A

decision making framework

Identify and define the problem: break problem down into modules Solve the problem: developing a model, collecting data, and analyzing data Perform Sensitivity Analysis: the model's predictions are investigated under alternative scenarios to see how our predictions/conclusions would change Communicate Results