Performance profiling of the unit trust funds in Malaysia with data mining techniques

Introduction

In Malaysia, different investment securities or schemes are publicly available to investors. The common securities include stocks, bonds, deposits, properties, unit trusts and commodities. Youngsters nowadays are more aware of the importance of earning passive income from investment. Bursa Malaysia recorded a 36% rise in the new Central Depository System (CDS) account holders aged below 25 in 2016, totalling more than 25.2 thousand account holders at the end of this year.¹ Recently, the fast-growing trend has driven more financial products to be brought to the market to attract investors.

Unit trusts are investment options with lower risk² than equities because they are well-diversified financial instruments that fund managers handle. Each unit trust fund has its investment objective and strategy. Hence an amateur needs to pick the suitable funds to be included in their portfolio. However, it is not easy to choose suitable unit trust funds that meet the investor’s requirements, especially from a large pool of unit trust funds available in Malaysia. As of December 2019, the Security Commission of Malaysia (www.sc.com.my/analytics/fund-management-products) reported that Malaysia’s total launched unit trust funds are 685.

Amateur investors always find it challenging to identify appropriate unit trust funds that meet their investment strategies and risks. The task requires much time and effort to complete, looking at the abundance of data to be searched and analysed. Most amateur investors are usually busy with their full-time jobs or running their own business. As a result, they cannot screen and perform in-depth analysis on many unit trust funds.

Another typically encountered problem is the low financial literacy among amateur investors. Thus, they cannot carry out a thorough analysis to extract meaningful information or knowledge from the annual reports they read. Subsequently, they find it challenging to analyse unit trust funds’ performance, and usually, their findings are inconclusive and indecisive. Amateur investors also face obstacles when they are looking for unit trust funds to fulfil their risk appetite. Hence, it is difficult for them to make investment decisions.

Considering the mentioned problems, this article aims to propose a data mining model that could build the performance profiles of unit trust funds using clustering and association rules mining algorithms. The model could mitigate the problems because: (i) the performance profiles could be generated rapidly, (ii) the financial information in the profile is intuitive and easy to understand by amateur investors, and (iii) the profiles also provide risk information.

Literature review

The research by F. Cai et al.³ evaluated 904 investment funds from financial datasets (time series and transactions) retrieved from the Morningstar using k-means and density-based clustering. The finding indicated that k-means performed better than the latter, giving the best number of clusters. In the research by T. Sakakibara et al.,⁴ clustering of mutual funds was done based on investment similarity instead of using historical performance similarity. The proposed approach was tested on 551 Japanese mutual funds. The result claimed that the approach could acquire the optimal number of clusters, even better than the classification provided by Morningstar Inc.

K. H. Ng et al.⁵ adopted association rules mining to find frequent financial patterns in outlier stocks listed on Bursa Malaysia. The outliers were identified via a score-based approach, and they were manually grouped into superior and poor outliers. The research produced nine rules associated with outstanding stocks and four rules for the poor performing stocks. Investors can refer to these rules for decision making. Another study applied clustering to construct an efficient stock portfolio in the Warsaw Stock Exchange.⁶ The study utilised k-means and Partitioning Around Medoids (PAM) methods for this purpose and showed satisfactory returns.

Methods

This research applies data mining techniques to profile the unit trust funds available in Malaysia. The entire process consisted of five key steps. The first step was to collect the financial data of the unit trust funds from annual reports. The next step involved a normalisation technique to ensure that the financial attributes share the same scale. The third step was grouping unit trust funds using a clustering algorithm. This was followed by profiling the clusters to find the common rules/characteristics associated with each cluster. In the final step, evaluation was conducted to study the efficiency of the proposed profiling method. The research overview is shown in Pseudocode 1.

Pseudocode 1. Profiling of unit trust funds.
Input: Annual financial reports of unit trust funds, AFRs
Output: Financial profiling of unit trust funds

1. //collect unit trust funds financial data
2. for each unit trust fund’s report, ut ϵ AFRs do
3.  for Fy ϵ {fromYr, toYr} do//financial year Fy
4.   FinD ← extract Fy data from ut
5.  //append financial data, FinD to dataset DbFy
6.   DbFy. append (FinD)
7.  end for
8. end for
9. //normalize the two financial datasets
10. perform z-score normalization on Db2017 and Db2018
11. //apply EM clustering to group unit trust funds
12. {C1, C2, C3, … , Cm} ← EMCluster (Db2017)
13. //EM generated m clusters, C1 until Cm
14. {E1, E2, E3, … , En} ← EMCluster (Db2018)
15. //EM generated n clusters, E1 until En
16. //discretize the attributes in the datasets
17. calculate quartile 1 – 3 values for all attributes in Db2017 and Db2018
18. for each attribute, att ϵ {Db2017, Db2018} do
19.   if att.value >= q3//quartile 3 value
20.    att.value ← high
21.   else if att.value > = q1
22.    att.value ← moderate
23.   else
24.    att.value ← low
25.   end if
16. end for
27. //find associated rules in each cluster
28. for each cluster generated in line 12 and 14
29.   apply apriori algorithm to find frequent itemset for the cluster
30.   profile the cluster based on the frequent itemset
31. end for

Pseudocode 2. EM clustering on the unit trust dataset.
Input: Unit trust dataset for the financial year, DbFY
Output: Unit trust clusters

1. //Frc represents the fraction of the data represented by cluster c
2. assign random values to mean0(c), variance0(c) and Fr0c
3. for each unit trust, utr in Dbyr do //expectation step
4.  for each cluster c do
5.   compute the probability of utr in the cluster c
6.  end for
7. end for
8. for each cluster c do//maximisation step
9.  re-estimate meani+1(c), variancei+1(c) and Fri+1c to maximise the likelihood of the unit trusts
10. end for
11. repeat steps 3 – 10 until the parameters converge
12. return the final clusters, c1, c2 … , ck //k denotes the number of clusters

Pseudocode 3. Mining rules associated with the cluster.
Input: Discretised financial data of unit trusts in a cluster
Output: Financial profiles of the unit trust cluster

1. //mSup denotes minimum support
2. L1 ← search all frequent 1-itemset with support count >= c_size * mSup //c_size is the cluster size
3. Initialize j ← 1
4. repeat
5.  j ← j+1
6.  CandISj ← find candidate itemsets from Lj-1
7.  for each unit trust utc in the cluster do
8.   Candutc ← subset (CandISj, utc)
9.   for each candidate itemset in Candutc do
10.    add support count of the candidate itemset
11.   end for
12.  end for
13.  Lj ← extract the frequent j-itemsets with support count >= c_size * mSup
14. until Lj is null
15. return all L

Results and discussion

This study used a unit trust dataset that contained six normalised financial variables of the year 2017. It started with the clustering process using the Expectation Maximisation (EM) method. As a result, eight clusters were produced. Table 2 shows the detail of each cluster, including the number of assigned unit trust funds and the means of the six financial features. Cluster 2 was the largest cluster with 63 unit trust funds, and the smallest was Cluster 5 with only 14 members.

Even though the mean of each financial variable is produced and displayed in the table, it is not an easy task for an amateur investor to analyse and interpret the generated information correctly. This is because the financial variables have been normalised before the clustering. Hence, a further step was taken to process this information by employing the Association Rules Mining technique so that the generated profile information on each cluster would be more intuitive and easier to understand. To perform the ARM, the continuous financial data was discretised into three bins, i.e., high, moderate, low, using the binning method.

ARM was conducted on each cluster to discover a set of rules associated with each cluster. Two parameters, i.e., minimum support and minimum confidence, were predefined. The first parameter was set to 50%,¹¹ and a higher threshold of 90% was applied to the minimum confidence. ARM only yielded meaningful outcomes on three clusters: Cluster 8, Cluster 4, and Cluster 5. The remaining clusters did not demonstrate any strongly associated rules after the process.

Cluster 8 produced two frequent 3-itemsets (Table 3). The result strongly indicates that unit trust funds in this cluster belong to the type of high total return, high capital growth, and high portfolio turnover. Upon examining the portfolios of some unit trust funds in the cluster, it was discovered that these funds prioritise stock investment in their investment baskets. This finding was sufficient to justify the derived rules associated with the cluster because the return and capital growth of stock investment were relatively higher than the fixed income securities such as bonds, fixed deposits, etc. High portfolio turnover is also expected in these unit trust funds because the fund managers must always fine-tune their portfolios to maximise the returns. As such, the financial profile of this cluster can be summarised as “Aggressive”. This cluster is appropriate for investors who are risk-takers and aim for a high growth rate and return.

The rules associated with Cluster 4 are illustrated in Table 4, displaying the three frequent 3-itemsets derived from the cluster. The financial profile of Cluster 4 might not have been as outstanding as Cluster 8, but it is noteworthy to describe the unit trust funds in Cluster 4 as average and Defensive. Even though they have average performance among their peers in total return, capital growth, and income distribution, they give stable returns over the years.

A thorough analysis of some of the unit trust funds in Cluster 4 revealed conservative investment approaches. They adopted a more diverse investment strategy by pooling their funds into a basket of securities consisting of shares and bonds, deposits, and properties. Combining these investment securities can lower the downside risk due to market volatility. In short, Cluster 4 was considered “Defensive”, and therefore suitable for investors with a low-risk tolerance.

Cluster 5 exhibited inferior financial performance. Table 5 shows the six frequent 2-itemsets derived after the ARM process. The associated rules in this cluster included low total return, low capital growth, high portfolio turnover, and high-income distribution. The phenomena of low total return but high-income distribution was due to poor capital growth. Stagnant or negative growth in the investment capital significantly outweighed the income distribution. Hence, this cluster was deemed as an inferior type. Investors are advised to avoid such unit trust funds at all costs.

The financial profiles created for the three clusters were further assessed and validated using the average total return (yearly) for the financial year 2018 and 2019 (Table 6). For 2018, the best performer was the Defensive cluster, which was the only one that still delivered a positive return (+1.17%). Both Aggressive and Inferior clusters were in the negative territory. Thorough investigation revealed that the unsatisfactory performance in 2018 was mainly attributed to the tumble of the Bursa market (Figure 1) affected by the three factors: general election, funds pulled out by foreign investors, and negative investor sentiment.

Figure 1. Bursa Malaysia KLCI Index lost more than 100 points in the year 2018.

The 14th Malaysia general election has shaken investors’ confidence to invest in Malaysia’s equity markets. As a result, many investors withdrew their securities investments, and their action caused a sharp fall in the Bursa Index.¹² The stock market was still in higher volatility mode in the second half of 2018. Many stocks could not recover their prices in the first half of the year.¹³ Funds pulled out by foreign fund managers also attributed to the stock market fall. It has been reported that the net foreign fund outflow in 2018 for Malaysia totalled RM11.65 billion.¹⁴ The last factor that adversely impacted the stock market was investor sentiment. Issues such as unresolved trade wars, lower oil prices, and geopolitical tension had refrained investors from securities investment.

In 2019, the volatility of the Malaysian stock market had subsided. Hence, all three clusters showed improvement in the total return as compared to the preceding year. The average total return (+4.65%) of Cluster 4 outperformed the others. Cluster 8 managed to gain a positive return of +3.52%. The worst performer (−2.98%) was still Cluster 5. The evaluation of two financial years for the three cluster profiles has strongly demonstrated that Cluster 4, with a Defensive profile, was resilient in the harsh economic climate that included market downturn, high market volatility, etc. This cluster may not have performed as well as Cluster 8 with an Aggressive profile during the economic boom, but it did provide steady passive income with lesser risk in the long run. Lastly, it could be summarised that investors with lower risk tolerance could aim at the defensive unit trust funds in Cluster 4. On the other hand, high-risk investors could identify the Aggressive unit trust funds in Cluster 8.

Conclusions

This study used data mining techniques to generate a financial profile for unit trust funds. The profile information is intuitive and easily understood by an investor. The investor could narrow down the number of unit trust funds for analysis. Investors could focus on just a single cluster for making an investment decision. For instance, the investor could identify aggressive unit trust funds from Cluster 8 if he/she is a high risk-taker.

On the contrary, investors could use Cluster 4 for steady and average passive incomes. The finding from this study will have significant implications for unit trust investors. To the best of our knowledge, no published work has applied data mining approaches to profile unit trust funds in Malaysia. We hope to continue and extend the study to include non-financial factors, e.g., fund manager profile and time-series data, in the future.

Data availability

Author contributions

All authors contributed equally to the conceptualisation of this study. Data curation, methodology, analysis and investigation, were performed by Aida Farah Khairudin. Keng-Hoong Ng and Kok-Chin Khor supervised the study and validated the evaluation results. Keng-Hoong Ng and Aida Farah Khairudin wrote the first draft of the manuscript. Kok-Chin Khor reviewed and edited the manuscript. All authors checked and approved the final manuscript.