forex_rnn.Rmd

---
title: "Forex Trading with RNN"
output:
  pdf_document: default
  word_document:
    fig_height: 4
    fig_width: 7
---


```{r setup, include=FALSE}
library(knitr)
knitr::opts_chunk$set(echo = TRUE, dpi=300, cache=TRUE, autodep=TRUE, cache.comments=FALSE,
               message=FALSE, warning=FALSE)
```

```{r imports, include=FALSE}
library(ggplot2)
library(dplyr)
library(lubridate)
library(keras)
library(wesanderson)
library(forecast)
library(TSstudio)
library(reshape2)
```

To simplify the problem, assume we always buy in at the start of the day and sell at the end of day. The key problem now is to predict days that will generate positive return. Assume we have $1000 principal at the start. 

# Read data into appropriate format
```{r read data}
# read the data while discard the last column
df <- read.csv('DAT_ASCII_GBPUSD_M1_2017.csv', sep = ";",
         col.names = c('timestamp', 'open', 'high', 'low', 'close','-'),
         stringsAsFactors=FALSE)[-6] %>%
  mutate(timestamp = as.POSIXct(timestamp, format="%Y%m%d %H%M%S")) %>%
  select(-c(high, low)) %>%
  mutate(date = date(timestamp)) 
```

# Data preprocessing
## Extract daily open and close price

```{r preprocessing}
daily.open <- df %>%
  group_by(date) %>%
  filter(timestamp == min(timestamp)) %>%
  ungroup() %>%
  select(open, date)

daily.close <- df %>%
  group_by(date) %>%
  filter(timestamp == max(timestamp)) %>%
  ungroup() %>%
  select(close, date)

daily.df <- daily.open %>% 
  merge(daily.close, by='date', 
        all.x=T, all.y=T) %>%
  mutate(month = month(date)) %>%
  mutate(return = close/open-1) %>%
  mutate(day_of_mon = mday(date))

summary(daily.df)
```

# Exploratory analysis 
## how does daily open and close price differ to each other?

```{r daily change}
melted <- daily.df %>%
  select(c(date, month, open, close)) %>%
  melt(id = c("date", "month")) %>%
  rename(type = variable, price = value) 

ggplot(data = melted, aes(x=date, y=price)) +
  geom_line(aes(color=type)) 
```

## What kind of daily return to expect?

```{r return}

daily.df %>%
  ggplot(aes(x=return))+
  geom_histogram(alpha=0.8, fill='#4682b4',bins=50)+
  geom_vline(xintercept = 0, color="red", linetype="dashed")+
  ggtitle("Histogram of Daily Price Movement") 
```

## Is there seasonality in return?

```{r season}
# seasonality of return 
daily.df %>%
  mutate(month = as.character(month)) %>% 
  ggplot(aes(x=day_of_mon, y=return) ) +
  geom_line(aes(color=month)) +
  geom_hline(yintercept = 0, color='black', linetype='dashed') +
  scale_color_manual(values=wes_palette(type = 'continuous', 18, 
                                        name = 'FantasticFox1')) +
  ggtitle('Daily Return in a Month')+
  xlab('Day of Month')+
  ylab('Daily Return')
```

# Trading Algorithm 

## Preparation: train-val split & function for backtest
```{r helper}
# train - val split 
train = daily.df[daily.df$month<=10,]
val = daily.df[daily.df$month>10,]

# helper function for backtest
BackTest <- function(decision_vec, principal=1000) {
  # args:
  #   1. decision_vec: vector of 1 or 0s indicating the days we are buying in 
  #   2. principal: be default to be 1000 
  # output:
  #   the profit 
  returns = val[decision_vec, 'return']
  profit = prod(returns+1)*principal - principal
  return(profit)
}
```

## A simple baseline
Baseline strategy: \\
buy if price went down on previou day; sell if price went up
```{r baseline model}
# baseline strategy: 
# buy if price went down on previou day; sell if price went up
decisions = c(train[nrow(train),'return'], val[1:nrow(val)-1, 'return'])
decisions = decisions >0

BackTest(decisions)
# -3.943526 
```
We will lose $3 if follow the baseline strategy. 

# Predict profitable days using recurrent NN 
## Prepare data 
```{r nn prepare}
y = daily.df$return>0
#one-hot-encoding
y_one_hot = to_categorical(y)

# normalise x
X = daily.df %>%
  select(-c(date, month, day_of_mon)) %>%
  scale()
X_array_expanded = array(0, dim = c(nrow(X), ncol(X), 1))

# create a test/validation set
X_array_expanded[,,1] = X

ndata = nrow(X)
n_train = as.integer(nrow(daily.df[daily.df$month<=10,]))
X_train = X_array_expanded[1:n_train,,]
dim(X_train) = c(n_train, ncol(X_train), 1)
X_valid = X_array_expanded[(n_train+1):ndata,,]
dim(X_valid) = c(ndata - n_train, ncol(X_train), 1)
```

```{r rnn model}
# RNN model 
input_X = layer_input(shape = c(ncol(X),1))
output_GRU_basic = input_X %>% 
  layer_gru(units=16, return_sequences = F,
            dropout=0.1) %>%
  layer_dense(units = 6, activation = "elu")  %>%
  layer_dense(units = 2, activation = "softmax") 
model_basic = keras_model(inputs = input_X, 
                          outputs = output_GRU_basic)

model_basic %>% summary()
```

```{r compile model}
model_basic %>% compile(
  optimizer = "rmsprop", 
  loss = "binary_crossentropy", 
  metrics = c("acc")
)

hist <- model_basic %>% fit(x=X_train,
                    y=y_one_hot[1:n_train,],
                    epochs = 60,
                    batch_size = 128,
                    validation_data = list(X_valid,
                                           y_one_hot[(n_train+1):ndata,]))
plot(hist)
```


```{r prediction}
#make predictions
library(ramify)  #for argmax
pred_proba = model_basic %>% predict(X_valid)
pred_class = argmax(pred_proba)

decisions <- pred_class - 1
BackTest(decisions)
```

We are making profit now. However, in the context of currency trading, false negatives are more detrimental to us, we shall adjust the threshold to make decision.

```{r adjust}
decisions.adjusted <- pred_proba[,2]-pred_proba[,1] >0.3
BackTest(decisions.adjusted)
```

We are now making `r BackTest(decisions.adjusted)/1000 * 100` % return.