Optimising the rsims package for Fast Backtesting in R – Part I

rsims is a new package for fast, quasi event-driven backtesting in R. You can find the source on GitHub, docs here, and an introductory blog post here.

Our use case for rsims was accurate but fast simulation of trading strategies. I’ve had a few questions about how I made the backtester as fast as it is – after all, it uses a giant for loop, and R is notoriously slow for such operations – so here’s a post about how I optimised rsims for speed.

Approach

I firstly wrote the backtester with a focus on ease of understanding. I wanted something that worked as expected, and that I could think about without too much effort. At this stage, I wasn’t thinking much about speed or efficiency.

Once I had that working, I focused on finding bottlenecks with profvis. It’s a good idea to resist the urge to optimise before you know where you’ll get the most bang for your buck (speaking from experience here).

Then I simply tackled those bottlenecks one at a time until I reached a point of diminishing returns (which is a little subjective and dependent on one’s use case).

There’s one thing I’d do differently if I were starting over: write the tests first, rather than after the fact. In hindsight, this would have saved a ton of time. To be honest, it’s something that I say after every little development effort, but this time I’ve really learned my lesson.

Profiling with profvis

The original code looked a lot different from the current version. It had lots of data frames and dplyr pipelines for operating on data:

positions_from_no_trade_buffer <- function(current_positions, current_prices, current_theo_weights, cap_equity, num_assets, trade_buffer) {
  current_weights <- current_positions*current_prices/cap_equity
  target_positions <- current_positions
  for(j in 1:num_assets) {
    if(is.na(current_theo_weights[j]) || current_theo_weights[j] == 0) {
      target_positions[j] <- 0
      next
    }
    # note: we haven't truncated to nearest whole coin, as coins are divisible (unlike shares)
    if(current_weights[j] < current_theo_weights[j] - trade_buffer) {
      target_positions[j] <- (current_theo_weights[j] - trade_buffer)*cap_equity/current_prices[j]
    } else if(current_weights[j] > current_theo_weights[j] + trade_buffer) {
      target_positions[j] <- (current_theo_weights[j] + trade_buffer)*cap_equity/current_prices[j]
    }
  }
  unlist(target_positions)
}
cash_backtest_original <- function(backtest_df_long, trade_buffer = 0., initial_cash = 10000, commission_pct = 0, capitalise_profits = FALSE) {
  # Create wide data frames 
  wide_prices <- backtest_df_long %>%
    pivot_wider(date, names_from = 'ticker', values_from = 'price')
  wide_theo_weights <- backtest_df_long %>%
    pivot_wider(date, names_from = 'ticker', values_from = 'theo_weight')
  # get tickers for later
  tickers <- colnames(wide_prices)[-1]
  # initial state
  num_assets <- ncol(wide_prices) - 1  # -1 for date column
  current_positions <- rep(0, num_assets)
  previous_theo_weights <- rep(0, num_assets)
  row_list <- list() 
  cash <- initial_cash
  # backtest loop
  for(i in 1:nrow(wide_prices)) {
    current_date <- wide_prices[i, 1] %>% pull() %>% as.Date()
    current_prices <- wide_prices[i, -1] %>% as.numeric()
    current_theo_weights <- wide_theo_weights[i, -1] %>% as.numeric()
    # update equity
    equity <- sum(current_positions * current_prices, na.rm = TRUE) + cash
    cap_equity <- ifelse(capitalise_profits, equity, min(initial_cash, equity))  # min reflects assumption that we don't top up strategy equity if in drawdown
    # update positions based on no-trade buffer
    target_positions <- positions_from_no_trade_buffer(current_positions, current_prices, current_theo_weights, cap_equity, num_assets, trade_buffer)
    # calculate position deltas, trade values and commissions
    trades <- target_positions - current_positions
    trade_value <- trades * current_prices
    commissions <- abs(trade_value) * commission_pct
    # adjust cash by value of trades
    cash <- cash - sum(trade_value, na.rm = TRUE) - sum(commissions, na.rm = TRUE)
    current_positions <- target_positions
    position_value <- current_positions * current_prices
    equity <- sum(position_value, na.rm = TRUE) + cash
    # Create data frame and add to list
    row_df <- data.frame(
       Ticker = c('Cash', tickers),
       Date = rep(current_date, num_assets + 1),
       Close = c(0, current_prices),
       Position = c(0, current_positions),
       Value = c(cash, position_value),
       Trades = c(-sum(trade_value), trades),
       TradeValue = c(-sum(trade_value), trade_value),
       Commission = c(0, commissions)
    )
    row_list[[i]] <- row_df
    previous_theo_weights <- current_theo_weights
  }
  # Combine list into dataframe
  bind_rows(row_list)
}

Stay tuned for the next part in which Kris will demonstrate how to make a data frame of randomly generated prices and weights.

Visit Robot Wealth to download the complete R script: https://robotwealth.com/optimising-the-rsims-package-for-fast-backtesting-in-r/.