FINNOMENA Best-In-Class Methodology White Paper
Once the Best not always the Best
To maximize the profit from investing in mutual funds, the individual investors normally select the mutual funds only based on the historical return of the given time period (e.g. 1 year, 5 years or 10 years). However, the historical data has shown that this strategy does not give the maximum profit compared to “FINNOMENA Best-in-Class”. In this paper, we explain the strategy behind our Best-in-Class fund selection using “deep learning”, one of machine learning methods which dominates the research in investment. We also show that our strategy using deep learning have significantly improved average alpha by 57.28% compared to the individual investors strategy.
Investing has become more popular and fashionable in Thailand recently. Among investment assets, mutual fund has been a solution of choice for individual investors who may not have time nor tools to analyze stock market. this has escalated the importance of mutual fund selection and portfolio construction method, which is the core of making profit in mutual funds. Investors often focus too much on the historical return of the given time period (e.g. 1-year return, 5-year return) as the main indicator of future fund performance. However, we found that portfolios constructed from this method does not yield a favorable result.
To enhance mutual fund portfolio return, we propose the new portfolio modeling method called FINNOMENA Best-In-Class (BIC). We have applied deep learning to help us select the best mutual funds in each category. Deep learning is one of machine learning methods which models a nonlinear relationship and often use for nonlinear forecasting. Our BIC model uses the massive historical fund data incorporate with deep learning to predict mutual funds future expected return, and form a BIC Portfolio. In the next section, we explore the strategy individual investors tend to rely on past performance and exposed the flaws of this method. Section 3 will explain the concept and reasoning behind the design of our deep learning model architecture. Lastly, we evaluate and show how our model performs in section 4.
2. Do Not Rely on Historical Return
Generally, individual investors tend to evaluate mutual funds based on historical return at the given period alone, which seems logical since historical return is in turn a reflection of the fund manager capability in managing their money. This causes individual investors to subconsciously make an assumption that funds which performed in the past will perform in the future as well; However, the following experiment give the evidence that this investment strategy is presumptuous hypothesis.
2.1 Data Source
In this experiment we use the historical NAV from Jan 2012 to Jan 2019, which will be explained further in Sec. 2.2, and chose to experiment on mutual funds in Equity Thai Active Large-Cap category, due to both its popularity and large sample space.
2.2 Portfolio Construction Method
The portfolio is formed and invested equally in three mutual funds that have generated the highest return in the past. And to replicate norm of the investor we chose three standard look-back periods (1, 3, and 5 years) as the parameter for calculating historical returns. This gave us 3 methods of portfolio construction using historical return.
2.3 Experiment Setup
In this experiment, we set the definition of the best portfolio to be the one that generates the maximum alpha in a given period of time.
α = ReturnPortfolio – ReturnBenchmark
However, instead of using market return as a benchmark, in this experiment we use the average return of mutual funds within the same category served as benchmark return to enhance our precision. Next, we evaluate the strategy based on the alpha that portfolio generated at the end of each of the given time periods. The invested period is set to be 6 months, meaning that each of the constructed portfolios will be invested for 6 months and the generated return during that period will be evaluated against the mutual funds’ return in the same category at the end of the 6th month as illustrated in Fig. 1. This experiment is repeated everyday start investing from January 2017 to June 2019, created the total of 549 samples, which portfolio returns were on, for each strategy.
Figure 1. Investment timeline
2.3 Experiment Results
we have found that the results are contrary to expectations of the norm. Intuitively, the funds that have proved themselves over a longer period of time should be more credible and therefore excel in the future as well.
Fig. 2 has shown that the portfolio constructed using 3 year historical return achieved the highest average alpha at 0.56%, while using 1 and 5 years historical return the average alpha were at 0.25% and -1.62% in order. On the other hand, investors would have got an average of 0% alpha if they were to invest at random.
Figure 2. Alpha distribution
3. Proposed Model
We propose a deep learning model that our quantitative analyst team has experimented on using the architecture shown in Fig. 3.2 to rank mutual funds within the same category, the reasoning behind each part will be further explained in Sec 3.2. Which on the contrary to the method in Sec 2.2, the ranking made by the model was based on the future expected return rather than the historical one.
3.1 Deep Learning
Figure 3.1 Deep Learning
Unlike traditional approaches such as ARIMA  that only rely on linear relationship between input data and output data to forecast time series data, deep learning models are much more effective at capturing complex structure and pattern in data using non-linear activation functions. Deep learning is a branch of machine learning; however, unlike it predecessors, deep learning is a natural inspired algorithm hoping to mimic how the human brain behaves. Similar to human brain that compose of neurons and synapses as shown in Fig. 3.1, deep learning processes input data in nodes (neurons) and pass on the information (signal) through edge (synapses) for the next layer.
3.2 Best In Class Model
Figure 3.2 Best-In-Class architecture
At the heart of this model, we use Recurrent Neural Network (RNN)  which is well known for its effectiveness and application in time series forecasting using internal memory that collects essential information from the past. This makes it a perfect choice for predicting time-dependent target. However due to the fact that the prediction made by a model has to be the same as label, this imposed a undesired restriction on the RNN memory dimension. So in order to remove this constraint, we apply a dense neural network simply to adjust the prediction dimension to be the same as the label.
On the downside, training a deep learning model takes tremendous amount of time, in order to mitigate this we improved the efficiency by incorporating a Batch Normalization  as a preprocessing method between each layer. This is aim to reduce the covariate shift, which occurs throughout the backpropagation process causing the delay in training process. By deploying this method, it does not only help accelerating the training process, but it allows us to use higher learning rate as well. This in turn reduces significant amount of resource required for training the model.
The error is then calculated using mean square error between model prediction and the label. We then use Adam Optimizer , which has proved to be state of the art gradient-based optimization in many literature, to compute and apply gradient throughout the model.
4. Best In Class Portfolio
After we got the score for each mutual fund from the model in Sec. 3, we then allocate the 3 highest scored mutual funds at 40%, 30%, and 30% in order to form a Best in Class Portfolio based on the chosen category. The portfolio is then adjust and rebalance every 6 months to ensure the performance as we have shown in Sec 2.4 that the historically performed mutual funds do not necessarily excel in the future. Please note that each Best In Class Portfolio used a specific model that was trained on the delegated category.
4.1 Best In Class Experiment Results
Using the same experimental method mentioned in Sec. 2.3, but this time instead of using the portfolio construction method as Sec. 2.2 we use the approach as mention in Sec. 4. The result has turned out to be better than what we anticipated. Comparing it to the result in Sec. 2.4, using the model has improved the average alpha to 0.89% shown in Fig 4., and not only that but it has a much smaller tail risks compare to method in Sec. 2.2.
Figure 4. Alpha Distribution with BIC
4.2 Trading Simulation
In order to further examine the model performance and understanding it limitation under the real world scenario, we have decided to test on rebalancing cycle parameters, every 3, 6, and 9 months, while fixed the switching period to be T+3, and transaction fee to be at 0.5% for both buy and sell, or 1% for each complete transaction. We have chosen two specific time periods to evaluate how well each of the rebalancing cycles performs during both rising and falling market environment compared with SET TRI shown in Fig. 5 and 6.
During the rising market, by rebalancing often the 3 months periods manages to pick up trend sooner than 6 and 9 months period even though it was heavily penalized by transactions fee but the reward was worth the cost when compared with SET TRI. However, rebalancing too often in falling market not only did it not help mitigate the fall but it was accelerated from the transaction fee. On the other hand, rebalancing every 9 months, which is the most cost-efficient approach in this case, was not affected much from the transaction fees, but it foregoes the opportunity of investing in quality mutual funds given by the model. Which leads us to rebalancing every 6 months method that performed well on both during the rising and falling market as the balance between cost and benefit of switching.
There are mainly two contributions we have made in this experiment. First, we have shown that using historical return alone as a method of portfolio construction is worse than random. Next, we have introduced a rather simple model, yet it allows the individual investor to have access to a strategy that has been back tested and evaluated on various condition, and most importantly we have successfully proved to outperform the typical strategy. We believed that the reason our model is able to achieve the average alpha at 0.89% is that the model was evaluating the fund manager performance overtime as part of the mutual fund’s score. Overall, we found this model to be well-suited for assisting individual investors to construct their portfolio.
- Alonso A.M., Garc ́ıa-Martos C., 2012. Time Series Analysis – Forecasting with ARIMA models. Universidad Carlos III de Madrid, Universidad Polit ́ecnica de Madrid.
- Sherstinsky, Alex. Fundamentals of Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) Network. ArXiv e-prints, November 2018.
- Ioffe, Sergey and Szegedy, Christian. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift. ArXiv e-prints, March 2015.
- Kingma, P. Diederik and Ba, Jimmy. Adam: A Method for Stochastic Optimization. ArXiv e-prints, January 2017.