generationk

⚠️ This is an early version. The package can be used for backtesting, but the API is not stable yet. This is a work in progress. Once stability is reached, version 1.0.0 will be tagged. Please check back then. ⚠️

Features

1. Read CVS file with historic data and back test in parallell. Testing ~5000 daily data per 100 CSV files can take about 1 sec. to complete.

ToDo

1. To set entry and exit conditions independently to be able to combine many different ones, ex. entry RSI < 20; exit MA2 > MA5
2. To be able to try which parameters are the best ones for an indicator (simplified branch)
3. Use a genetic algo to scan the search space for most profitable parameters
4. Optimize for other functions than profit, drawdown, volatility
5. Make a WS API, deploy as lambda functions
6. It would be interesting to output some common statistics about back test except profit etc.

generationk

The inspiration for this project took place after using a few other backtesting frameworks in Python.

I was tired of waiting for results and concluded that I want the fast feeling of a compiled language and I wanted all processor cores to be used.

I looked at a few different ones in Golang but they where either very complex for simple tasks or did not really appeal to the context.

Design choices

I have created 3 versions of this framework.

1. The very first version was based on channels and real time but it very rarely stock data is real time in backtesting, even a minute is considered very granular and 5 minutes or 10 minutes are much more common. In that case reading a 2-3 data points will advance time 20-30 minutes just for placing an order which is not realistic in the end.

2. The second version was based on callbacks using interfaces which would have the advantage of feeding one data point after the other into the backtester from the simulated data source and this way obtain a more realistic feeling. A main argument is that it is not possible to peek at future data by mistake. In the end I think a backtest that peeks into future data stands out in the results as amazing and is actually easy to spot. I also felt that everything gets very slow and the reason is not strong enough since backtests are not realistic anyway. My main priority is to check statistics to see if there is an edge and get a quick overview and then to see how I could manually trade a system.

3. In the third version which is the current what is provided is basically a for loop which loops over data. I have also tried to remove all kinds of encapsulation and special 'types'. I try to as much as possible only work with float64. Because it gives a degree of freedom and any functions can be written to work on numbers. So I want to as much as possible have access to the raw data and numbers instead of a 'LinearNumberSeries' struct and so on...

The Crossing MA example looks like this

type MACrossStrategy struct {
	ma50  []float64
	close []float64
}

//Once is used to declare what indicators will be used; it's run once
func (ma *MACrossStrategy) Once(ctx *K.Context, ohlc *K.OHLC) error {

	//The closing prices
	ma.close = ohlc.Close

	//The Simple Moving Average length 50 periods, the ones from 0 to 50 
	//will be registred in the array as well
	ma.ma50 = indicators.SimpleMovingAverage(ohlc.Close, 50)

	//If the init period is set, the PerBar function below will not be called 
	//until the InitPeriod is reached
	ctx.SetInitPeriod(50)

	//No errors to my knowledge
	return nil
}

//PerBar gets called when there is new data coming in
func (ma *MACrossStrategy) PerBar(k int, callback K.Callback) error {

	//Closing price breaks the MA50
	if ma.close[k] > ma.ma50[k] {
		//Are we owning this stock since before?
		if !callback.Owning() {
			//No. Then we can buy 100 stocks with a marketorder
			callback.SendOrder(K.BuyOrder, K.MarketOrder, 100)

to run the above strategy a command like this could be used:

go run generationk/cmd backtest -test MACrossStrategy -dir ../test/data/CSV2/ -fromDate 01/01/2015

Strategies

I want the strategies directory to contain working trading strategies that are actually used in the market. There are many working ones and the hard part is to trade the system, actually not find a system that is profitable. If you find something interesting, please share it. Sharing is caring 🤗

Future implementation

• To use genetic algorithms to find the best trading system for a stocks / stocks
• To plug in machine learning

Literature on backtesting

Even though we back test we are still biased. In their work on representativeness, Tversky and Kahneman (1974) find that people expect that a sequence of outcomes generated by a random process will resemble the essential characteristics of that process even when the sequence is short. Griffin and Tversky (1992) provide an extension documenting that people focus on the strength or extremeness of the evidence with insufficient regard of its credence, predictability, and weight.