Implenent Parabolic SAR - Extended (SAREXT)

[AlexCatarino]

Expected Behavior

Lean support the Parabolic SAR - Extended (SAREXT) Indicator.

Actual Behavior

Not supported. See #83.

Potential Solution

Implement SAREXT. There is scarce information on this indicator. We could look into TALIB's formulas to understand the difference from PSAR.

Checklist

• [x] I have completely filled out this template
• [x] I have confirmed that this issue exists on the current master branch
• [x] I have confirmed that this is not a duplicate issue by searching issues
• [x] I have provided detailed steps to reproduce the issue

[LouisSzeto]

spy_sarext.csv

[nbkell]

I would like to be assigned this as a first issue if it hasn't been handled yet. @AlexCatarino could you assign it to me? Thanks!

[nbkell]

@AlexCatarino I’ve recently made a PR that implements SAREXT, which is simply a generalization of PSAR (called SAR in talib) that gives you more control over the parameters and behavior. The differences are:

• SAREXT is negative if you’re in a short position, whereas PSAR is always positive regardless of current position.

• You can now specify different acceleration factor parameters (start, increment, and max) that used according to whether you’re in a short or long position.

• You can now specify a starting value of the indicator (where the sign of the starting value dictates whether the indicator starts in a long or short position).

• An additional parameter called offsetOnReverse, which determines a percentage of the current SAR that is first incremented/decremented from the indicator before switching between long/short positions.

  Comments/questions regarding testing:

• To create my test .txt I just used the applied SAREXT function from tallib to spy ticker data I obtained from yahoo finance. I’ve attached the ticker data file and Python script I used to make the test file below.

• I’d be happy to add the file posted by @LouisSzeto, but I’m not sure how to incorporate it if doesn’t include values for low, high, and close?

Finally I want to point out that the current Lean implementation for PSAR is not entirely consistent with that of SAR in talib (see the link to ta_SAR.c below). These discrepancies are:

• talib does not consider the first bar when initializing the extreme point _ep.
• talib uses the DM_MINUS function to determine if indicator starts in long/short position (currently LEAN compares close direction of first and second bars).
• In the samples == 2 case, ComputeNextValue() should still do the first iteration of long/short updates (instead of just returning _sar computed in init()).

I’ve implemented SAREXT to be consistent with talib (and thus some of the differences from Lean’s PSAR). Let me know if you’d rather it be consistent with Lean’s current version of PSAR.

Files and References:

• Ticker data: SPY.csv
• Python script to generate test file: make_sarext_test.zip
• talib SAR code
• talib SAREXT code

[LouisSzeto]

Hi please refer to this reference data for checking: spy_sarext.csv

The columns are "DATE" "SPY high" "SPY low" "SAREXT"

The script generating the data

import talib
import pandas as pd

history = pd.read_csv("https://github.com/QuantConnect/Lean/raw/master/Data/equity/usa/daily/spy.zip",
                       index_col=0, names=["open", "high", "low", "close", "volume"])

high = history.high
low = history.low

sarext = talib.SAREXT(high, low).dropna().to_frame()
sarext["high"] = high
sarext["low"] = low
sarext = sarext.iloc[:, [1, 2, 0]]
sarext.to_csv("spy_sarext.csv", header=False)

[jaredbroad]

Welcome to LEAN @nbkell!