time series - Poor prediction using ARIMA model

I am trying to fit and forecast log returns of a price data using ARIMA model in R. For reproducibility, data is provided here.

Steps Followed, Code and Results obtained

1. Check for outliers (Package: forecast) - No outliers detected.

outliers <- tsoutliers(log.rtn)

2. Stationarity Check using ADF test (Package: fUnitRoots) - Series found to be stationary

stationary <- adfTest(log.rtn, lags = m1\$order, type = c("c"))

3. Determination of p,d,q using ACF and PACF (Package: astsa) - Based on my understanding, p = 2, d = 0, q = 2

acf2(log.rtn, lags = 20)

4. Fitting ARIMA (Package: forecast)

fit <- auto.arima(log.rtn, stepwise=FALSE, trace=TRUE, approximation=FALSE)


Model obtained : ARIMA(2,0,1)

Series: log.rtn

ARIMA(2,0,1) with zero mean

Coefficients:
ar1     ar2     ma1
-0.5705  0.1557  0.6025
s.e.   0.1549  0.0532  0.1519

sigma^2 estimated as 0.001086:  log likelihood=775.57
AIC=-1543.14   AICc=-1543.04   BIC=-1527.29

5. Prediction (Package:forecast)

fcast <- forecast(fit, n.ahead=5)
plot(fcast)

Point Forecast       Lo 80      Hi 80       Lo 95      Hi 95
390   1.416920e-03 -0.04080849 0.04364233 -0.06316127 0.06599511
391   8.228924e-04 -0.04142414 0.04306993 -0.06378837 0.06543416
392  -2.488236e-04 -0.04289257 0.04239493 -0.06546681 0.06496917
393   2.700663e-04 -0.04248622 0.04302635 -0.06512003 0.06566016
394  -1.928045e-04 -0.04303250 0.04264690 -0.06571047 0.06532486
395   1.520366e-04 -0.04273465 0.04303872 -0.06543749 0.06574156
396  -1.167506e-04 -0.04303183 0.04279833 -0.06574971 0.06551621
397   9.027370e-05 -0.04284167 0.04302221 -0.06556846 0.06574901
398  -6.967566e-05 -0.04301167 0.04287232 -0.06574379 0.06560444
399   5.380284e-05 -0.04289419 0.04300179 -0.06562948 0.06573708


I am quite confused why the model is predicting so badly.

• What makes you think that the predictions are poor? Jul 13, 2015 at 13:10
• @forecaster: When I plot the data used for fitting (389 data points) and predictions obtained. The predicted values to be almost constant. Apart from that, the magnitudes of the series and predicted values are different.. Jul 13, 2015 at 13:25
• That tells you the data is patternless and has low forecastability, things like price follow random walk and cannot be forecasted, use theory and judgement rather than relying on an automated algorithms. Jul 13, 2015 at 14:55
• @forecaster: Thanks for your suggestion. So, it is always advisable to first plot the series before proceeding for model building. I would also like to know whether the sequence of approach applied is correct, otherwise, to apply on some other data? Jul 13, 2015 at 15:37

For log returns the recommended model to use is GARCH and its variations. Log returns are characterised by volatility clusters: periods of high volatility are followed by high volatility and periods of low volatility are followed by low volatility.

GARCH is designed to handle volatility in a much better way than ARIMA. Further I would not treat the data for outliers as a perceived outlier could carry signal on the start (or end) of a volatility cluster.

Check this post, the R package fGarch and the function garch from package tseries.

• While this link may answer the question, it is better to include the essential parts of the answer here and provide the link for reference. Link-only answers can become invalid if the linked page changes. Jul 24, 2015 at 0:07
• The answer is: poor predictions are a result from the algorithm adopted. GARCH would provide better predictions as it deals better with log returns data. The link serves as additional documentation on GARCH. In the unlikely event of link changes, a simple Google search on GARCH would be equally useful. Jul 30, 2015 at 6:52
• Nice edits. (My comment and delete vote were made prior.) Jul 30, 2015 at 13:11
• @AbhayBhadani, note that using a GARCH model is not likely to help much in getting accurate forecasts of the conditional mean of the series (even though it might be a good model for the conditional variance). The conditional mean of stock returns, commodity returns, futures returns and many other financial series is nearly unpredictable, and there is little you can do about it besides accepting it. Aug 3, 2015 at 17:58

There are a few mistakes on your reasoning. Let me explain

• The Augmented Dickey–Fuller test is NOT a stationarity test. The null-hypothesis on this test is that there exists a unit-root in the series. Rejecting this hypothesis DOES NOT mean the series is stationary. It just means that there is enough evidence to allow you rejecting the hypothesis of a very specific form of non-stationarity (i.e. unit root)
• To see an ARMA is not a good model for your data, do estimate the ACF of the squares of log-returns. You will see significant autocorrelation at high lags. Which means there is high memory in volatility and/or volatility is non-stationary

• There is plenty of evidence in financial literature to support the idea that volatility is clustered and that log-returns are uncorrelated. When trying to fit a model, you should not discard prior evidence and theory (when available). This should be your starting point