Automated Trading - Train, Test, Forecast
ContextBase, contextbase.github.io
Programming by John Akwei, ECMp ERMp Data Scientist
March 20, 2017
Synopsis
The objective of this document is to apply training and testing datasets, (from Machine Learning techniques), to stock trading analysis in the R programming language. This document utilizes the “QuantMod” and “PerformanceAnalytics” R packages for Backtesting of Automated Trading Stategies. Holt-Winters and GARCH Filtering are used for Price Prediction.
Working Directory, and Required Packages
# Set Working Directory
setwd("C:/Users/johnakwei/Dropbox/Programming/StockTrading/YahooFinanceGainersLosers")
# Required Packages
# if (!require("quantmod")) { install.packages("quantmod"); require("quantmod") }
# if (!require("PerformanceAnalytics")) { install.packages("PerformanceAnalytics"); require("PerformanceAnalytics") }
# if (!require("rgl")) { install.packages("rgl"); require("rgl") }
# if (!require("rugarch")) { install.packages("rugarch"); require("rugarch") }
library(ggplot2)
library(xts)
library(zoo)
library(tseries)
library(forecast)
library(quantmod)
library(PerformanceAnalytics)
library(caret)
library(rugarch)Downloading Stock Ticker Data from Yahoo Finances
# Import Original data
ticker_lookup <- getSymbols("GOOG", auto.assign=F,
from="2014-01-01", to="2017-01-01")
# write.csv(ticker_lookup, "ticker_lookup.csv")
ticker_lookup2 <- as.data.frame(cbind(ticker_lookup[,1], ticker_lookup[,2], ticker_lookup[,3], ticker_lookup[,4], ticker_lookup[,5], ticker_lookup[,6]))
TCKR <- as.data.frame(cbind(date = rownames(ticker_lookup2), ticker_lookup2))
colnames(TCKR) <- c("date","open","high ","low ","close ","volume ","adjusted")
rownames(TCKR) <- NULL
rm(ticker_lookup, ticker_lookup2)Create Training and Testing data sets
# Create Training and Test Data
TCKRTrain <- TCKR[1:567,]
TCKRTest <- TCKR[568:756,]Create Time Series data
# Time Series Processing
TCKRTrainTS <- ts(TCKRTrain$close, start=c(2014, 1, 1),
end=c(2016, 4, 5), frequency=250)
TCKRTestTS <- ts(TCKRTest$close, start=c(2016, 4, 4),
end=c(2017, 1, 1), frequency=250)Forecasting Model #1 - Holt Winters Filtering
# Fit Exponential Smoothing
fit_ES <- HoltWinters(TCKRTrainTS)
# Forecast 20 Days
TCKRForecast <- forecast(fit_ES, h=20)
# Accuracy of Exponential Smoothing Predictive Model
accuracy(TCKRForecast, TCKRTestTS)## ME RMSE MAE MPE MAPE MASE
## Training set 0.7567337 12.36518 6.315612 0.1584173 1.090240 0.01061984
## Test set -33.0843421 49.40416 35.392489 -4.6871517 4.991725 0.16992610
## ACF1 Theil's U
## Training set -0.005877338 NA
## Test set 0.848029852 4.549523# Realized Data Plot
plot(TCKR$date, TCKR$close, col="orange", xlab="Date",
ylab="Prices", main="Plot of Ticker Data")
# Exponential Smoothing Plot
plot(TCKRForecast, xlab="Time", ylab="Prices",
main="Holt-Winters Forecast - 20 Days")
# Output Exponential Smoothing Data to Excel
TCKRFrcstSeries <- c(TCKRTrain$close[1:567], TCKRForecast$mean[1:20])
# .csv of original data and all 3 forecasted values
TCKRMerge <- c(TCKRTrain$close[1:567], rep(NA, 20))
dateColumn <- c(as.Date(TCKRTrain$date[1:567]),
as.Date(TCKRTest$date[1:20]))
TCKRMergeTest <- c(rep(NA, 567), as.numeric(TCKRTest$close[1:20]))
allData <- data.frame(TCKRMerge, TCKRFrcstSeries)
Comparison <- data.frame(dateColumn, TCKRMerge,
TCKRMergeTest, TCKRFrcstSeries)
names(Comparison) <- c("Date", "Train_Data", "Test_Data", "HW_Forecast")
tail(Comparison, 30)## Date Train_Data Test_Data HW_Forecast
## 558 2016-03-21 742.09 NA 742.0900
## 559 2016-03-22 740.75 NA 740.7500
## 560 2016-03-23 738.06 NA 738.0600
## 561 2016-03-24 735.30 NA 735.3000
## 562 2016-03-28 733.53 NA 733.5300
## 563 2016-03-29 744.77 NA 744.7700
## 564 2016-03-30 750.53 NA 750.5300
## 565 2016-03-31 744.95 NA 744.9500
## 566 2016-04-01 749.91 NA 749.9100
## 567 2016-04-04 745.29 NA 745.2900
## 568 2016-04-05 NA 737.80 747.7955
## 569 2016-04-06 NA 745.69 747.7115
## 570 2016-04-07 NA 740.28 750.0822
## 571 2016-04-08 NA 739.15 744.3641
## 572 2016-04-11 NA 736.10 741.4964
## 573 2016-04-12 NA 743.09 745.8130
## 574 2016-04-13 NA 751.72 751.1398
## 575 2016-04-14 NA 753.20 752.6830
## 576 2016-04-15 NA 759.00 759.4063
## 577 2016-04-18 NA 766.61 758.4920
## 578 2016-04-19 NA 753.93 769.9277
## 579 2016-04-20 NA 752.67 786.5731
## 580 2016-04-21 NA 759.14 792.4501
## 581 2016-04-22 NA 718.77 788.0976
## 582 2016-04-25 NA 723.15 771.9331
## 583 2016-04-26 NA 708.14 763.7044
## 584 2016-04-27 NA 705.84 764.7288
## 585 2016-04-28 NA 691.02 788.9019
## 586 2016-04-29 NA 693.01 783.1890
## 587 2016-05-02 NA 698.21 784.2773plot(Comparison$HW_Forecast[1:587], type="l", col="blue",
main="Forecast Comparison Plot", xlab="Time", ylab="Values")
lines(Comparison$Test_Data[1:587], col="green")
lines(Comparison$Train_Datat[1:587], col="red")
# write.csv(allData, "allTickerData.csv")
# write.csv(Comparison, "TickerDataComparison.csv")Forecasting Model #2 - Holt Winters Exponential Smoothing
tickerData.xts <- xts(as.numeric(Cl(TCKRTrain)),
order.by=as.Date(TCKRTrain$date))
tickerData.z = zoo(x=Cl(TCKRTrain),
order.by=as.Date(TCKRTrain$date))
# Specify the prices and store our models
prices <- tickerData.xts[,1]
# Create indicator
sma <- SMA(tickerData.xts, n=1)
# Calculate the indicators we need for our strategy
CCI20 <- CCI(prices, 20)
RSI3 <- RSI(prices, 3)
DEMA10 <- DEMA(prices, n = 10, v = 1, wilder = FALSE)
DEMA10c <- prices - DEMA10
DEMA10c <- DEMA10c/.0001
buy.signal <- ifelse(RSI3 < 30 & CCI20 > -290 & CCI20 < -100 & DEMA10c > -40 & DEMA10c < 750, 1, NA)
# Construct trading rule
sig <- Lag(ifelse(sma$SMA < buy.signal, 1, -1))
# The trading rules/equity curve
retSMA <- ROC(tickerData.xts) * sig
chartSeries(tickerData.xts, theme = chartTheme('white'),
TA=c(addVo(),addBBands(), addMACD()))
TCKRTS <- ts(TCKRTrain$close, start=c(2014, 1, 1),
end=c(2016, 1, 1), frequency=250)
# Seasonal decomposition
fit <- stl(TCKRTS, s.window="period")
monthplot(TCKRTS, col="green", main="Month Plot of TWTR, Seasonal Adjustment")
# triple exponential - models level, trend, and seasonal components
fit <- HoltWinters(TCKRTS)
# predict next three future values
TWTRForecast <- forecast(fit, 20)
plot(TWTRForecast, main="TWTR Price Forecast (20 Days) from Holt-Winters Filtering")
# New Forecast Data
TWTRFrcstSeries <- c(TCKRTrain$close[1:567], TWTRForecast$mean[1:20])
Comparison$NewForecast <- TWTRFrcstSeries
names(Comparison) <- c("Date", "Train_Data","Test_Data",
"HW_Forecast", "New_Forecast")
tail(Comparison, 30)## Date Train_Data Test_Data HW_Forecast New_Forecast
## 558 2016-03-21 742.09 NA 742.0900 742.0900
## 559 2016-03-22 740.75 NA 740.7500 740.7500
## 560 2016-03-23 738.06 NA 738.0600 738.0600
## 561 2016-03-24 735.30 NA 735.3000 735.3000
## 562 2016-03-28 733.53 NA 733.5300 733.5300
## 563 2016-03-29 744.77 NA 744.7700 744.7700
## 564 2016-03-30 750.53 NA 750.5300 750.5300
## 565 2016-03-31 744.95 NA 744.9500 744.9500
## 566 2016-04-01 749.91 NA 749.9100 749.9100
## 567 2016-04-04 745.29 NA 745.2900 745.2900
## 568 2016-04-05 NA 737.80 747.7955 758.6653
## 569 2016-04-06 NA 745.69 747.7115 757.7018
## 570 2016-04-07 NA 740.28 750.0822 747.3589
## 571 2016-04-08 NA 739.15 744.3641 736.0305
## 572 2016-04-11 NA 736.10 741.4964 735.7642
## 573 2016-04-12 NA 743.09 745.8130 737.9528
## 574 2016-04-13 NA 751.72 751.1398 732.0523
## 575 2016-04-14 NA 753.20 752.6830 729.0013
## 576 2016-04-15 NA 759.00 759.4063 733.1340
## 577 2016-04-18 NA 766.61 758.4920 738.2772
## 578 2016-04-19 NA 753.93 769.9277 739.6377
## 579 2016-04-20 NA 752.67 786.5731 746.1743
## 580 2016-04-21 NA 759.14 792.4501 745.0739
## 581 2016-04-22 NA 718.77 788.0976 756.3278
## 582 2016-04-25 NA 723.15 771.9331 772.7907
## 583 2016-04-26 NA 708.14 763.7044 778.4858
## 584 2016-04-27 NA 705.84 764.7288 773.9525
## 585 2016-04-28 NA 691.02 788.9019 757.6062
## 586 2016-04-29 NA 693.01 783.1890 749.1947
## 587 2016-05-02 NA 698.21 784.2773 750.0360plot(Comparison$HW_Forecast[1:587], type="l", col="blue",
main="Forecast Comparison Plot", xlab="Time", ylab="Values")
lines(Comparison$New_Forecast[1:587], col="black")
lines(Comparison$Test_Data[1:587], col="green")
lines(Comparison$Train_Datat[1:587], col="red")
Forecasting Model #3 - GARCH (Generalized AutoRegression of Conditional Heteroskadacity) Forecasting
spec <- ugarchspec()
nrow(expand.grid(GARCH=1:14, VEX=0:1, VT=0:1, Mean=0:1, ARCHM=0:2,
ARFIMA=0:1, MEX=0:1, DISTR=1:10))## [1] 13440spec <- ugarchspec(variance.model=list(model='eGARCH', garchOrder=c(1, 1)),
distribution='std')
# Fit models with Generalized Auto-Regressive Conditional Heteroskadacity
all.fitted.model <- ugarchfit(spec, TCKRTrainTS, solver='hybrid')
coefBMtable <- data.frame(coef(all.fitted.model))
names(coefBMtable) <- "GARCH Parameters"
coefBMtable## GARCH Parameters
## mu 1.113141e+03
## ar1 9.996090e-01
## ma1 -1.872695e-03
## omega 5.864394e-02
## alpha1 3.870138e-02
## beta1 9.871125e-01
## gamma1 3.092318e-02
## shape 3.189047e+00par(mfrow = c(1,2))
plot(all.fitted.model, which=8)
plot(all.fitted.model, which=9)
par(mfrow = c(1,1))
forc1 = ugarchforecast(all.fitted.model, n.ahead=20)
plot(forc1, which = 3)
# GARCH Forecast Data
GARCHFrcstSeries <- c(TCKRTrain$close[1:567],
forc1@forecast$seriesFor[1:20])
Comparison$GARCHForecast <- GARCHFrcstSeries
names(Comparison) <- c("Date", "Train_Data", "Test_Data", "HW_Forecast",
"New_Forecast", "GARCH_Forecast")
tail(data.frame(Comparison[1], Comparison[3:6]), 30)## Date Test_Data HW_Forecast New_Forecast GARCH_Forecast
## 558 2016-03-21 NA 742.0900 742.0900 742.0900
## 559 2016-03-22 NA 740.7500 740.7500 740.7500
## 560 2016-03-23 NA 738.0600 738.0600 738.0600
## 561 2016-03-24 NA 735.3000 735.3000 735.3000
## 562 2016-03-28 NA 733.5300 733.5300 733.5300
## 563 2016-03-29 NA 744.7700 744.7700 744.7700
## 564 2016-03-30 NA 750.5300 750.5300 750.5300
## 565 2016-03-31 NA 744.9500 744.9500 744.9500
## 566 2016-04-01 NA 749.9100 749.9100 749.9100
## 567 2016-04-04 NA 745.2900 745.2900 745.2900
## 568 2016-04-05 737.80 747.7955 758.6653 759.0415
## 569 2016-04-06 745.69 747.7115 757.7018 759.1799
## 570 2016-04-07 740.28 750.0822 747.3589 759.3183
## 571 2016-04-08 739.15 744.3641 736.0305 759.4567
## 572 2016-04-11 736.10 741.4964 735.7642 759.5950
## 573 2016-04-12 743.09 745.8130 737.9528 759.7332
## 574 2016-04-13 751.72 751.1398 732.0523 759.8714
## 575 2016-04-14 753.20 752.6830 729.0013 760.0095
## 576 2016-04-15 759.00 759.4063 733.1340 760.1476
## 577 2016-04-18 766.61 758.4920 738.2772 760.2856
## 578 2016-04-19 753.93 769.9277 739.6377 760.4236
## 579 2016-04-20 752.67 786.5731 746.1743 760.5615
## 580 2016-04-21 759.14 792.4501 745.0739 760.6993
## 581 2016-04-22 718.77 788.0976 756.3278 760.8371
## 582 2016-04-25 723.15 771.9331 772.7907 760.9749
## 583 2016-04-26 708.14 763.7044 778.4858 761.1126
## 584 2016-04-27 705.84 764.7288 773.9525 761.2502
## 585 2016-04-28 691.02 788.9019 757.6062 761.3878
## 586 2016-04-29 693.01 783.1890 749.1947 761.5254
## 587 2016-05-02 698.21 784.2773 750.0360 761.6628plot(Comparison$HW_Forecast[1:587], type="l", col="blue",
main="Forecast Comparison Plot", xlab="Time", ylab="Values")
lines(Comparison$GARCH_Forecast[1:587], col="orange")
lines(Comparison$New_Forecast[1:587], col="black")
lines(Comparison$Test_Data[1:587], col="green")
lines(Comparison$Train_Datat[1:587], col="red")