Put the FUN in Mutual Funds

Research Questions

1. Can we group open-end mutual funds based on sales patterns in the last 2 years?

Motivation: If I, asset manager x, can find funds similar to my own which are not currently captured by Morningstar Category relationships, I can target those funds for competition.

2. Can we segment broker dealer (BD) offices into groups with similar buying behavior?

Motivation: I, asset manager x, can better arm my salesforce to pursue opportunities if we have targeted strategies for each segment of the market.

Data: Open End Mutual Fund Sales aka What was sold, when, and where?

• When: Monthly aggregates January 2016 - December 2017
• Where: Broker Dealer Locations
• What: CUSIP, Fund ID, Morningstar Category

---

Fund Clustering

Time Series

• Hypothesis: funds that are similar have similar sales patterns.
• Features: monthly sales into each FundID in the rolling 24 months
• Distance Metric: We want to know how closely the sales movements of these funds align. Dynamic Time warp can help us measure this.

Dynamic Time Warp

Dynamic time warp is an algorithm to measure similarity between two temporal sequences, which may vary in speed.

I used an implementation written by Pierre Rouanet: https://github.com/pierre-rouanet/dtw

Dynamic Time Warp for sin vs sin: 0.0
Dynamic Time Warp for sin vs cos: 0.04
Dynamic Time Warp for sin vs sin*2: 0.28
Dynamic Time Warp for sin vs sin+2: 1.0

Hierarchical Clustering

Well-suited to situations where want to understand the relationships within clusters. For example, within this cluster, which funds are the most similar? Are there any funds that are total snowflakes?

Interpretation:

• Most funds are pretty similar by this metric
• There are some funds that are definite outliers - really far away from the rest of the pack.

---

Broker Dealer Clustering

• **Hypothesis: Broker Dealers/Broker Dealer Offices with similar buying behavior can constitute meaningful consumer segments. **
• Features: Proportion of sales into each broad category within the broker dealer or broker dealer location
  • Included scaled broker dealer size for the broker-dealer level analysis because size is a meaningful differentiator for sales opportunity.

Broker Dealer Office Clustering by Broad Category

Ameriprise

Ameriprise Cluster Centroids

Allocation	Alternative	Commodities	Convertibles	Equity	Fixed Income	Tax Preferred
7%	3%	0%	0%	28%	56%	6%
8%	3%	0%	0%	56%	27%	6%
29%	6%	0%	0%	27%	25%	12%

Broker Dealer Clustering by Broad Category

BD Clusters overlayed on PCA	BD Clusters overlayed on t-SNE

Broker Dealer Cluster Centroids

Allocation	Alternative	Commodities	Convertibles	Equity	Fixed Income	Tax Preferred	BD Size vs Largest
45%	0%	0%	0%	36%	13%	5%	0%
8%	6%	1%	0%	19%	17%	48%	0%
12%	0%	0%	0%	72%	11%	4%	0%
9%	1%	0%	0%	14%	72%	5%	0%
17%	2%	0%	0%	42%	32%	7%	1%
11%	2%	0%	0%	36%	40%	10%	68%

Continued Exploration:

FundID

• Scale the fundid data to make all time series the same magnitude.
• Add in redemptions and returns features.
• Feature engineer a "target" to represent which funds actually do sell together at the same location or unseat one another.

Broker Dealer + Broker Dealer Offices

• Try category instead of broad category.
• Build deep neural net to predict sales into categories within each broker dealer office for the next month/quarter.

Resources

• dtw module by pierre-rouanet https://github.com/pierre-rouanet/dtw python -m pip install dtw