The Premise

We want to predict the performance of Ethereum against USD using historical data. Generally, this practice is frowned upon because each new trading day changes the business cycles that produce following rates. However, for the sake of this assignment, instead of using regressors, we are going to try to create a model based on categorization and feature engineering.

Our data is a bit rough, with only daily readings of Open, High, Low, Close, Volume, (OHLC) and Block Size. Finer-grain data is a bit expensive.

Instead of predicting an exact exchange rate, we will simply try to determine if the following week from any data point is bullish (significantly increasing) or bearish (significantly decreasing).

library(tidyverse)
library(zoo)
library(smooth)
library(TTR)
library(tidymodels)
library(plotly)

data <- "https://raw.githubusercontent.com/TheWerefriend/data607Final/main/ETH-USD.csv" %>%
  read.csv(header = TRUE) %>%
  tibble() %>%
  mutate(Date = as.Date(Date, format = "%Y-%m-%d"))

data <- data[, -6] %>%
  mutate(Open = as.numeric(as.character(Open))) %>%
  mutate(High = as.numeric(as.character(High))) %>%
  mutate(Low = as.numeric(as.character(Low))) %>%
  mutate(Close = as.numeric(as.character(Close))) %>%
  mutate(Volume = as.numeric(as.character(Volume))) %>%
  na.locf()

# na.locf() from zoo replaces NA with previous non-NA value

Feature engineering

We will be using volume to establish a feature called Momentum, which is the rate of change of the volume of trades. This will tell us if there is significant investor emotion regarding this asset.

data <- data %>%
  arrange(Date) %>%
  mutate(Momentum = ROC(Volume, n = 7, type = "continuous"))

The next feature will be MACD (Moving Average Convergence/Divergence), which shows differences between exponential moving averages of two different periods. If momentum is significant, this feature will tell us what direction the price is moving.

fast <- 7
slow <- 30

data <- data %>%
  arrange(Date) %>%
  mutate(SMA12 = sma(Close, h = fast)$fitted[,1])

MACD <- MACD(data$Close, nFast = fast, nSlow = slow)

data <- cbind(data, MACD)

Next, we will include a Stochastic oscillator, which gives the relationship between the asset’s closing price and the price range over the previous period.

stochastic <- stoch(data$Close, nFastK = 14,
                    nFastD = 3, nSlowD = 3)

data <- cbind(data, stochastic)

Finally, we must look into the future to score the Outlook for each observation as bullish or not.

data <- data %>%
  arrange(Date) %>%
  mutate(Future = as.factor(ifelse(
    lead(Close, n = 7) - Close > 0,
    TRUE, FALSE)))

We now must remove the trailing NAs we introduced with out lagged/leading indicators.

data <- na.omit(data)

The Model

We will split the data at 2/3 and train a model (probably random forest) with independent variables Momentum, MACD, Stochastic, and dependent being the Outlook.

split <- initial_split(data, strata = Future, p = 0.67)
trainer <- training(split)
tester <- testing(split)

ethRecipe <- recipe(Future ~ Momentum + macd +
                      signal + fastK + fastD + slowD,
                    data = trainer) %>%
  step_center(all_predictors()) %>%
  step_scale(all_predictors()) %>%
  prep(training = trainer)

ethTrain <- juice(ethRecipe)
ethTest <- bake(ethRecipe, tester)

ethModel <- rand_forest(mode = "classification") %>%
  set_engine("ranger")

ethFit <- ethModel %>%
  fit(Future ~ Momentum + macd + signal + fastK + 
        fastD + slowD, data = tester)

Scoring the Model

Precision? Recall? F-measure?

results <- ethTest %>%
  select(Future) %>%
  mutate(predicted = factor(ifelse(ethFit$fit$predictions[,1] > 0.6, TRUE, FALSE)))

precision(results, truth = Future, estimate = predicted)
## # A tibble: 1 x 3
##   .metric   .estimator .estimate
##   <chr>     <chr>          <dbl>
## 1 precision binary         0.430
recall(results, truth = Future, estimate = predicted)
## # A tibble: 1 x 3
##   .metric .estimator .estimate
##   <chr>   <chr>          <dbl>
## 1 recall  binary         0.758
f_meas(results, truth = Future, estimate = predicted)
## # A tibble: 1 x 3
##   .metric .estimator .estimate
##   <chr>   <chr>          <dbl>
## 1 f_meas  binary         0.549

Visualizations

Use the levels function to determine the positive and negative classes. Here we see that FALSE is the positive class.

levels(data[["Future"]])
## [1] "FALSE" "TRUE"

Confusion matrix

Confusion Matrix - Our model correctly classified 220 out of 332 predictions (66.3%). Which translates into: - 126 correct FALSE predictions - 87 correct True predications

conf_mat(results,
         truth = Future,
         estimate = predicted)
##           Truth
## Prediction FALSE TRUE
##      FALSE   229  304
##      TRUE     73   57

Plot the confusion matrix as a heatmap for our model.

 conf_mat(results,
         truth = Future,
         estimate = predicted) %>%
  autoplot(type = 'heatmap')

Plot the confusion matrix as a mosaic for our model. Here the height of the FALSE - FALSE combination represents the sensitivity wile the height of the TRUE - TRUE combination represents the specificity.

conf_mat(results,
         truth = Future,
         estimate = predicted) %>%
  autoplot(type = 'mosaic')

eth_class_preds <- ethFit %>%
  predict(new_data = tester,
          type = 'class')

eth_prob_preds <- ethFit %>%
  predict(new_data = tester,
          type = 'prob')

eth_results_tester <- tester %>%
  select(Future) %>%
  bind_cols(eth_class_preds, eth_prob_preds)

The default probability threshold in binary classification is 0.5. The positive class in this model is FALSE as shown by the levels() function above. This means that if the .pred_FALSE >= 0.5 then .pred_class is set to “FALSE” by the predict() function.

Plot the receiver operating characteristic (ROC) curve which we use to visualize performance across a range of probability thresholds. Ideally, we would like to see the points as close to the upper left edge of the plot (0,1) as possible. This would indicate a strong performance of our model. However, one can see here that the points lie close the diagonal were sensitivity equals specificity thus indicating a strong performing model.

eth_results_tester %>%
  roc_curve(truth = Future, .pred_FALSE) %>% 
  autoplot()

We can summarize the ROC curve by calculating the area under the curve (AUC). In this case, the ROC AUC is 0.999 which translates to an A in the letter grade scale below.

roc_auc(eth_results_tester,
        truth = Future, .pred_FALSE)
## # A tibble: 1 x 3
##   .metric .estimator .estimate
##   <chr>   <chr>          <dbl>
## 1 roc_auc binary          1.00

By using plotly, an interactice plotting library, it’s possible to chart financial data such as ours.

Candlestick Chart

Source: https://plotly.com/r/candlestick-charts/

data_cs <- as.data.frame(data)
#cs <- data_cs %>% plot_ly(x = ~Date, type = "candlestick",
#                          open = ~Open, close = ~Close,
#                          high = ~High, low = ~Low)
#cs <- cs %>% layout(title = "Near all-time $Eth Candlestick Chart")
#cs
data_cs <- data.frame(data_idx=index(data_cs),coredata(data_cs))

# create Bollinger Bands
bbands <- BBands(data_cs[,c("High","Low","Close")])

# join and subset data
data_cs <- subset(cbind(data_cs, data.frame(bbands[,1:3])), Date >= "2015-08-06")

# colors column for increasing and decreasing
for (i in 1:length(data_cs[,1])) {
  if (data_cs$Close[i] >= data_cs$Open[i]) {
      data_cs$direction[i] = 'Increasing'
  } else {
      data_cs$direction[i] = 'Decreasing'
  }
}

increase_color <- list(line = list(color = '#17BECF'))
decrease_color <- list(line = list(color = '#7F7F7F'))
cs_chart <- data_cs %>% plot_ly(x = ~Date, type="candlestick",
          open = ~Open, close = ~Close,
          high = ~High, low = ~Low, name = "ETH",
          increasing = increase_color, decreasing = decrease_color)
cs_chart <- cs_chart %>% add_lines(x = ~Date, y = ~dn, name = "B Bands",
            line = list(color = '#ccc', width = 0.5),
            legendgroup = "Bollinger Bands", inherit = F,
            showlegend = FALSE, hoverinfo = "none") 
cs_chart <- cs_chart %>% add_lines(x = ~Date, y = ~mavg, name = "Mv Avg",
            line = list(color = '#E377C2', width = 0.5),
            hoverinfo = "none", inherit = F)
cs_chart <- cs_chart %>% layout(yaxis = list(title = "Price"))

# plot volume bar chart
cs_chart_2 <- data_cs 
cs_chart_2 <- cs_chart_2 %>% plot_ly(x=~Date, y=~Volume, type='bar', name = "ETH Volume",
          color = ~direction, colors = c('#17BECF','#7F7F7F')) 
cs_chart_2 <- cs_chart_2 %>% layout(yaxis = list(title = "Volume"))
# create rangeselector buttons
range_sel <- list(visible = TRUE, x = 0.5, y = -0.055,
           xanchor = 'center', yref = 'paper',
           font = list(size = 9),
           buttons = list(
             list(count=1,
                  label='RESET',
                  step='all'),
             list(count=1,
                  label='1 YR',
                  step='year',
                  stepmode='backward'),
             list(count=3,
                  label='3 MO',
                  step='month',
                  stepmode='backward'),
             list(count=1,
                  label='1 MO',
                  step='month',
                  stepmode='backward')
           ))
# subplot with shared x axis
cs_chart <- subplot(cs_chart, cs_chart_2, heights = c(0.7,0.2), nrows=2,
             shareX = TRUE, titleY = TRUE)
cs_chart <- cs_chart %>% layout(title = paste("ETH: 2015-09-01 -",Sys.Date()),
         xaxis = list(rangeselector = range_sel),
         legend = list(orientation = 'h', x = 0.5, y = 1,
                       xanchor = 'center', yref = 'paper',
                       font = list(size = 10),
                       bgcolor = 'transparent'))

cs_chart

Conclusions

This model is absolutely useless for predicting Ethereum. It seems, no combination of changes in parameters, models, or hyperparameters can solve this issue. What we have stumbled upon is the age-old wisdom that you simply cannot use historical timeseries data to predict new data points. Don’t try. Each new piece of data represents changes to the global process which creates the asset’s movement. There is no established cycle, there is no reason to believe that past performance indicates future performance.

While these forms of technical analysis could work for well-establish stocks or bonds (to a very limited extent), there is a strong need for leading indicators that do not come from the timeseries data.