Recency, Frequency and Intensity (RFI)
Introduction
Recency, frequency and intensity (RFI), is an analytical technique typically used in customer relationship management (CRM) (but not exclusive to it). It aims to understand the behaviour or responsiveness [8] of customers within a rolling time period [9]. Depending on the granularity of the time dimension, recency is the temporal measurement of how recently an interaction has occurred [9, 4, 2]. Frequency measures how often interactions occur within the same time period. Intensity measures how productive the interactions have been [9, 4, 2]. Within CRM, intensity is commonly measured monetarily leading to the variation RFM [9, 4, 2].
A Segmentation Approach
RFI takes a segmentation approach to categorising customer behaviour. Hughes suggests segmenting customers by first splitting recency into quintiles [8]. The top twenty percent of customers, ordered by most to least recent, are given a recency score of five. The following twenty percent are awarded a four and so on. This is repeated for frequency and intensity. Each customer is then assigned a combined RFI score. Miglautsch highlights the flexibility of this method by adding weights to each metric depending on its perceived importance [11]. However, this segmentation approach fails to “take advantage of all available customer data and not simply RFM counter variables” [6].
Data Mining and Clustering
An alternative to quintile segmentation is to use data mining to unearth “valid, useful, and understandable” patterns [2]. It aims to answer questions posed by the business, for example how can customers be segmented or which are the most valuable customers [1]. It can find associations between data for example the likelihood of purchase of a product following the purchase of another [1]. Data mining applies descriptive and predictive analytical algorithms such as the K-Mean algorithm [3], Sequential Pattern Mining (SPM) [4] or the CHAID algorithm [9] to data. It segments customers by finding natural meaningful behavioural clusters from trends using the RFI metrics [9]. The discovered clusters are subsequently tagged.
Tagging
Tagging is the process of hierarchically classifying the meaning of identified clusters once the centroids (mean) [3] have been established. A tag will contain as alias together with a description. A customer is assigned a tag based on the cluster they appear in for the period analysed. Tagging allows non-technical users to understand how RFI analysis has classified the customer.
Dimensional Storage
Muntean suggests storing the RFI metrics and tag in the customer dimensions table [12]. If the business is only interested in the current state of customers, this data can be stored as a type 1 slowly changing dimension (SCD). However, it is likely the business will want to analyse customer behaviour over time, and identify how customers migrate from one cluster to another in response to marketing campaigns or seasonal behaviours. Kimball views tags as “facts summarising behaviour” [9]. He recommends storing tags for a set number of time periods in columns in the dimension table using a type 3 SCD. In comparison to using a fact table or a type 2 SCD, less data is generated and comparing tags across time periods is easier [9]. Eventually the number of columns available will be exhausted requiring more columns to be added, or creation of a new row [9]. There is a potential issue if there is a change in granularity, as this will change the periodic meaning of the tag columns. It could be argued based on the granularity of time whether these dimensions are rapidly changing [13]. Excessively wide rows can lead to poor performance so it is prudent to think about the alternative of using a mini-dimension table (type 4 SCD) to store the historical tag values. The current tag dimension will therefore be of type 1 SCD. This solution is a good compromise between the amount of data stored and with the fact table holding a foreign key to the mini dimension table, good performance for querying across periodic tag values.
Advantages and Disadvantages
RFI analysis is not geared towards making predictions unlike statistical models which aim to make predictions on, for example, those customers that will respond to direct marketing. What it can do is provide a business with the means to gain competitive advantage through an understanding of customer behaviour. For example, it can define patterns of similar customer behaviour within clusters and migration from one cluster to another [1]. The tagging of these clusters allows terabytes of data can be boiled down into simple tags [9] for storage.
Businesses already collect and store the data required for RFI analysis. The model is simple, intuitive and cost-effective [11]. It can be used as a base to build on by adding additional metrics and weighting. Wei et al argue that RFI is effective “as the purchase behaviour can be summarised by using a very small number of variables” [15].
Wei et al criticises RFI segmentation which by its very nature focuses on “the best customers” [15]. It can potentially miss the untapped potential of recent customers, who have made a single small purchase. The segmentation approach is also likely to result in a large bucket of 1-1-1 customers, but unlikely statistical analysis, it cannot predict the probability of response to marketing for these customers [5].
The alternative use of Data Mining provides the power to use as many metrics as collected in analysis. Although RFI metrics can be extended to include demographic and other captured attributes, Ye acknowledges that it fails to understand critical purchasing triggers “such as taste/brand preferences or price sensitivity” [16]. It also fails to distinguish between the number of purchases made and the “weight of importance” [16] of the items, for example between multiple apples and one television.
Although the model can make predictions, these will be based on historical data. It is unable to highlight prospective new customers or potential target markets [5].
Applications of RFI
The RFI analytical technique is applicable to a range of areas outside of CRM.
For example, RFI can find the most unresponsive or responsive constituents to direct mailing, whether this is for product advertising, fundraising or monetary donors. This information can then be used for tailored and targeted marketing.
Within the military, studies have used the variation Recency, Frequency, Duration (RFD) to measure the link between combat exposure and alcohol abuse. Within this study, the metrics were extended to include age, marital status and enlistment type. The outcome however found RFI to have limitations within its results, but deemed worthwhile due to it highlighting requirements for a more detailed study [14].
Healthcare spending is a prime example of how RFI analysis can be used effectively to highlight patients with the highest regular expenditure who could be targeted for intervention to minimise long term costs. Recency measures how recently the patient has required treatment. Frequency is the number of treatments within the time period measured. Intensity measures the expenditure over this period. Data mining can use this data to find the clusters of patients who are current, regular and expensive where intervention, such as private treatment, would be recommended. This study can be modified to use RFD to analyse patients with particular ailments that may require regular stays in hospital. An alternative to hospitalisation, such as home care schemes, can be crucial in minimising costs [7]. Data mining can be used to highlight those patients where this intervention can be recommended.
Conclusion
RFI has the ability to highlight behaviours and trends to institutions that would not otherwise be found.
It is not as powerful as more complex statistical modelling tools which are geared towards prediction [5]. In terms of CRM, RFI analysis has the ability to provide competitive advantage to a business through simple customer analysis, providing an understanding of current and past customer behaviour.
RFI has been shown to be a useful tool. It can even be used to compliment predictive models [3].
References
[1] Buttle, F. Customer Relationship Management, 2nd Ed, Elsevier, 2009. Pg 114-117
[2] Chao, P. Fu, H. Lee, H and Chang Y. Identifying The Customer Profiles for 3C-Product Retailers: A Data Mining Approach, International Journal of Electronic Business Management, 2008, Vol. 6, No. 4, pp. 198-202.
[3] Cheng, C. and Chen, Y. Classifying the segmentation of customer value via RFM model and RS Theory, Expert Systems with Applications Journal, No. 36, 2009.
[4] Chen, Y. Kuo, M. Wu, S. and Tang, K. Discovering Recency, Frequency, and Monetary (RFM) Sequential Patterns From Customers’ Purchasing Data, Electronic Commence Research and Applications Journal, No. 8, 2009.
[5] D’Auria, T. The Fall of RFM Analysis, Business Intelligence Unlocked, June 2009.
[6] Drozdenko, R. G. And Drake, P. D. Optimal Database Marketing: Strategy, Development, and Data Mining, SAGE, 2002
[7] Escarrabill, J. Discharge Planning and Home Care for End-Stage COPD Patients, European Respiratory Journal, Vol. 34, Issue 2, pp. 507-512, August 2009
[8] Hughes, A. M. Strategic Database Marketing, 2nd Edition, McGraw-Hill, 2000
[9] Kimball, R. Wrangling Behaviour Tags, Intelligent Enterprise, 2002.
[10] McCarty, J. A and Hastak, M. Segmentation Approaches in Data-Mining: A Comparison of RFM, CHAID and Logistic Regression, Journal of Business Research, Vol. 60, 2007 pp. 656-662.
[11] Miglautsch, J. Thoughts on RFM Scoring, The Journal of Data Marketing, Vol. 8, No. 1, August, 2000.
[12] Muntean, O. Data Warehouse Solutions for CRM, International Conference on Computer Systems and Technologies (CompSysTech), 2004
[13] Ponniah, P. Data Warehousing Fundamentals for IT Professionals, John Wiley and Sons, 2010, pp. 249- 261
[14] Spera, C. Relationship of Military Deployment Recency, Frequency, Duration, and Combat Exposure to Alcohol Use in the Air Force, Journal of Studies on Alcohol and Drugs, Vol 72, Issue 1, pp 5(10), January 2011
[15] Wei, J. Lin, S. and Wu, H. A Review of the Application of RFM Model, African Journal of Business Management, Vol. 4 (19) December 2010
[16] Ye, N. The Handbook of Data Mining, Lawrence Erlbaum Associates, 2003, Pg 259-261
Related posts:
