Here is in Algorithm platform.

Glad you brought this up! There is a subtle but often overlooked issue (source of frustration, headache, annoyance) with exponentially weighted moving averages. Namely, the window length is sometimes not given. Take for example Wilders definition

WMA = price today * K + EMA yesterday * (1-K)

There isn't an explicit indication of how many 'yesterdays' to go back in the recursive calculation. It's good practice to define the window length as well as the weighting factor. If different window lengths are used then different results will happen. That is part of the issue why the notebook and the algo values, and associated graphs, are different. The window lengths are different.

Specifically there are three issues why the two two are different. Here's the short version...

1. The notebook had no initial lookback data while the algo did (the algo starts with 240 days of additional lookback).
   
2. The notebook used an 'expanding' window of lookback data (keeping the start date fixed) while the algo had a fixed length 'rolling window' of lookback.
   
3. The notebook used SPY pricing which was adjusted as of the end_date. The algo used data which was adjusted as of each simulation day (similar to pipeline data).

That's the summary. For a step by step outline of the issues and how to get the two aligned take a look at the attached notebook.

Hi Dan,

Thanks for your great explaination!

But truely to say I find the smooth one ist more reallity than the zick-zack one. I use many other chart tolls and they all show the smooth one.

How to can I get the smooth one in algo?

I find out, the longer I set in hist = data.history(context.spy, 'close', context.ma_length + 40, '1d'), the smoother it will be. But the question is: If the underlying is emited on 2010.01.01 for example and I do the backtesting from 2010.02.01, I can get so many historical datas and there for the WMA will be wrong, right?

Besides, even I use the very long back datas in the hist(), the WMA looks still have some small zick-zack on Dec.19, Sep.19, June 19 etc.(see attached backtesting below). You wrote in your notebook:
... The notebook used SPY pricing which was adjusted as of the end_date. The algo used data which was adjusted as of each simulation day (similar to pipeline data).
...

Is it possible to make the pricing in algo adjusted as of the end_date?

Thomas

@Thomas,

Wilder's MA in terms of digital signal processing is an infinite impulse response filter (IIR).
Its value today substantially depends of starting point of data.
To make it smooth and mature in IDE just increase number of data points 4-8 times.

...

@Vladimir

You are right. I also notice I have to set the data points very long to get a smoother MA. But eventhough, there is still some very small zick-zack on the date of Dec.19, Sep.19, June 19 etc.

you can also see these in your attached algo if you zoom in it. In notebook there is no such zick-zack. I wonder this might be caused from the non-adjusted close in algo?

Thomas

@Thomas,

These small spikes may be the result of Quantopian’s sophisticated way to apply adjustments for splits, dividends, and other distributions.
Try it on instruments that do not pay dividends.

@Vladmir

Thanks!

Since we can't do the live trade by Q, I hope these "spikes" will not happen by other trading platforms. :-)

@Thomas,

Quantopian applies adjustments for splits, dividends, and other distributions in a different way than most of us are used to.
Don't worry about these visual little splashes on Wilder's MA.
The Quantopian engine, when you apply trading logic or technical indicators behind the scene treats it like a smooth line.

Here is a test that proves this.
There are no sell transactions.

@Vladmir

Many thanks again!