Chasing Returns of Open-End Investment Funds Using Recurrent Neural Networks. A Long-Term Study

Figure 1.

Figure 2.

Literature review on forecasting open-end investment fund NAV and performance using machine learning (chronological order)

Author(s)	Year	Prediction objective	Dataset employed	Frequency of data	Machine learning prediction method	Error measures	Overall results
Chiang, Urban, Baldridge	(1996)	NAV p.s.	1981–1986 101 US mutual funds	5 years to predict year 6	BPN vs regression models	MAPE	BPN model provided better predictions compared to regression models based on MAPE
Indro et al.	(1999)	1-factor Jensen's alpha	1993–1995 559 US equity funds	3 years (1 year to predict 1 year)	MLP with GRG2	ME, MAE, MAPE, MSE	MLP model outperformed other models based on multiple error measures
Lin et al.	(2007)	NAV p.s.	3 single national equity funds of Taiwan, US and Japan		RBFNN	Error Index (EI)	RBFNN found effective
Wang and Huang	(2010)	Sharpe index	3 historical periods 1995–2000 Mutual funds listed in the Taiwan Economic Journal	72 months (1 year to predict 1 year; every two years)	FANNC vs BPN	RMSE	FANNC model outperformed the BPN in terms of RMSE, providing more accurate predictions
Yan et al.	(2010)	NAV p.s.	1 equity Chinese investment fund		BPN		good prediction accuracy
Ray and Vina	(2011)	NAV p.s.	1999–2004 10 funds from India	60 months	BPN		BPN demonstrated strong performance in predicting fund values
Priyadarshini and Babu	(2012)	NAV p.s.	2003–2009 1 fund	84 months	BPN	MAE, MSE, RMSE, MAPE, MPE	Error measures indicate solid performance of the BNP model
Priyadarshini	(2015)	NAV p.s.	2006–2012 1 fund	72 months	MLP	MAE, MSE, MAPE, RMSE, MPE	Good predictive performance based on these error metrics
Narula, Jha, Panda	(2015)	NAV p.s.	15-Oct-2012 till 2-Jan-2014; 200 Indian funds	300 consecutive trading days	FLANN vs RBF vs MLP	MAPE	FLANN performed well according to MAPE
Anish and Majhi	(2016)	NAV p.s.			RBF and FLANN	MAPE, RMSE	both models performed well, with FLANN having a slight advantage in terms of MAPE and RMSE
Anish, B. Majhi, R. Majhi	(2018)	NAV p.s.			RBF-PSO in comparison to MLANN, FLANN and RBFNN	MAPE, RMSE	The RBF-PSO model was the most accurate according to MAPE and RMSE
Han et al.	(2018)	NAV p.s.	31-Aug-2015 till 1-Jul-2016 2 funds	210 days	GRNN	RMSE, RTIC, MAE, MAPE, CE	GRNN provides highly accurate predictions
Pan et al.	2019	NAV p.s.	31-Aug-2015 till 1-Jul-2016 17 balanced open-end funds	210 days	BPN vs GABPN vs multiple regression	RMSE, RTIC, MAE, MAPE, CE	BPN model showed superior performance
Das et al.	(2020)		SBI Magnum Equity and UTI Equity	2010	BPN, RBPNN, RRBFNN	MSE, RMSE, MAPE	RBPNN outperformed over the other two prediction methods
Rout, Koudjonou, Satapathy	(2020)	NAV p.s. (normalized)	1998–2002 5 equity funds	1065–1255 days (80% of days in training and 20% of days in testing)	FLANN	RMSE, MAPE	FLANN found effective
Li and Rossi	(2020)	Carhart (1997) 4-factor adjusted alpha	1980–2018 2980 US equity funds	10 years of training to predict 1 subsequent year alpha	BRT, lasso, elastic net, random forest, NN	MAE, MSE, RMSE	Especially BRT and random forest outperform traditional regression models in predicting fund performance
Kaniel et al.	(2023)	4/5/6/8-factor Jensen's alpha	1980–2019 3275 U.S. equity funds	last month or year data to predict the next month	FFN	MAE	FFN models provided accurate predictions
DeMiguel et al.	(2023)	6-factor Jensen's alpha	1980–2020 8767 US equity funds	10 years of training to predict 1 year alpha	Gradient boosting: random forest, elastic net	MAE, MSE, RMSE	these advanced machine learning models performed well in prediction accuracy over long training periods.

Returns of the strategy and its benchmarks

	average return	return for the best fund	return for the worst fund
All funds
Strategy based on RNN fund return predictions	34.12%	245,76%	−29,67%
ARIMA model	30.88%	241.03%	−28,65%
“buy and hold” strategy	30.44%	241.03%	−30.22%
Equity funds
Strategy based on RNN fund return predictions	33.74%	85.21%	−28.88%
ARIMA model	29.14%	84.36%	−31.22%
“buy and hold” strategy	29.47%	84.36%	−30.22%
Hybrid funds
Strategy based on RNN fund return predictions	36.29%	246.54%	−16.95%
ARIMA model	32.12%	242.37%	−17.41%
“buy and hold” strategy	31.96%	241.03%	−17.68%
Fixed-income funds
Strategy based on RNN fund return predictions	33.25%	56.21%	17.03%
ARIMA model	31.01%	54.90%	16.49%
“buy and hold” strategy	31.58%	54.90%	16.49%
Money market funds
Strategy based on RNN fund return predictions	27.21%	38.12%	20.12%
ARIMA model	24.98%	37.51%	18.77%
“buy and hold” strategy	25.09%	37.51%	19.88%

Descriptive statistics of studied funds (annualised logarithmic rate of returns)

	No. of funds	Avg NAV (in Mio PLN)	logarithmic return
			max	min	avg	median	st.dev.	1st quartile	3rd quartile
all	71	769.5	775%	−447%	4.1%	3.9%	11.9%	−7.1%	16.8%
equity funds	18	602.72	443%	−447%	3.9%	4.6%	19.6%	−30.7%	43%
hybrid funds	24	714.35	775%	−283%	4.7%	5.2%	10.7%	−11.1%	22.1%
fixed-income funds	21	658.46	231%	−92%	4.2%	3.8%	4.9%	−1%	9.6%
money market fund	8	906.68	44%	−70%	3.8%	3.6%	1.5%	2.2%	5.2%