New Challenge: Build Smart Beta Factors

Algorithm template:

Disclaimer

Factor NB template:

Disclaimer

Andy Y

I like the long only strats idea because after the coronavirus blows over, there is no where but up for the market - just as it was after 2008. Citadel went on a hiring spree in 2009 to build out their teams because anything you could throw your money at would go up in that period of time. Also, the questionare is currently blocked and one cannot submit their answers to it

@Thomas,

Regarding turnover less than 10% monthly, are we allowed to use the MaxTurnover constraint in the optimizer to control this if needed?

@ Thomas,

why are the total returns in backtester and alpha decay analysis notebook so different?

@Vladimir: In the example I posted? I might have used a different backtest ID.

Disclaimer

@ Thomas,
Look at my notebook above.
This is my backtest and my notebook on that backtest.

@Vladimir, @Thomas,

I have the same observations as Vladimir. I originally thought that the returns on the notebook were annualized but appparently not. It is the same for other metrics such as turnover and Sharpe (IR), they are different from what the full tearsheet in Backtest IDE shows.

Oh yeah, that looks wrong. Let me investigate.

Disclaimer

@Vladimir @James: Could you try with this NB?

Disclaimer

@Thomas, seems like cummulative returns and IR is now consistent with backtest results. I'm not sure about monthly turnover . My backtest shows a daily turnover of 0.785% (or 0.785 x 21days = 16.485%), the graph on the NB shows much higher monthly turnover.

Here's my revised Model 1:

Nadeem Ahmed

First Entry ever :-) - Model 1
Pure Smart Beta.

Just to confirm, the turnover graph is monthly turnover (rather than daily) and is written as a decimal, correct? So 0.10 on the graph = 10% monthly turnover, not 0.1% monthly turnover, yes?

@ Thomas,

I have tried new notebook.
The returns and sharpe metrics very consistent with backtest results.

Arun

Model #1

Arun

Model #2

Arun

Model #3

@All: I just updated the above alpha-decay NB as it contained a bug and gave you wrong results that did not match those with what your backtester showed (the backtest results were more accurate). I also relabeled the lower-left plot as that is showing Monthly turnover to make it easier to spot whether you breach the 10% monthly turnover limit.

Disclaimer

Model 1

@James, @Thomas,

I have the same observations as James:
My backtest shows a daily turnover of 0.785% (or 0.785% x 21days = 16.485%),
the graph on the NB shows much higher monthly turnover.

With scheduled rebalance once a year I see in my notebook monthly turnover at 25% level between trades
which may be translated as 350% a year and not realistic.
Backtest shows 0.15% daily turnover which may be translated as 3.15% a month or 37.8% a year what is more likely correct.

May be we need to change this line

monthly_turnover = factor_stats['turnover'].resample('1M').sum()

@Thomas, I’ve observed the same as James and Vladimir.

Arun

I'm seeing the same thing too

Arun

Model #4

Arun

Model #5

Arun

Model #6

Georg Keller

Model 1

My first submission :)

I'm investigating the turnover questions that came up. Here is a NB. The difference is that the alpha decay analysis uses the positions to compute turnover while pyfolio uses transactions. The two should come out in a similar way, and they do (although I'm not sure if in my NB the difference is still larger than expected). Could someone try if the difference is much larger for them?

Disclaimer

Georg Keller

Model 2

@Thomas, here's the result of the monthly turnover test based on my Model 4. In this case, I see substantial difference between pyfolio and position based monthly turnover. If transactions based pyfolio monthly turnover was used, it would have probably made the below 10% threshold.

@ Thomas,

Here is 13 years backtest of simple buy and hold strategy.
Turnover in backtaster 0%.

@ Thomas,

Here's your last notebook of the same backtest.
The positions change sometimes reaches 40% and never reaches 0%.

@Thomas, @Vladimir,

I think the disconnect between pyfolio which is transactions based vs. Thomas' NB which is positions based is the actual execution of trades based on rebalancing schedule. Positions can change on a daily basis but when you schedule to rebalance on a monthly or quarterly basis, actual execution of trades takes place as scheduled. Therefore, the NB is overstating turnover as it is recording turnover as positions change but not actually executed. Makes sense?

Great demonstration of the turnover problem @Vladimir!

@James Yes, you're right, just the regular stock movements due to price shifts will increase the position-based turnover calculation. I suppose the easiest fix would be to change the alpha decay NB here to compute turnover based on transactions, we do have that data available anyhow. Anyone can think of a better solution?

Disclaimer

@Thomas, I concur. I think it is the most straight forward fix.

@all: I updated the top NB to use a more accurate turnover calculation. It should now be lower than what you've seen before.

I'm also attaching a version of the template algorithm that uses downsample() to only update a factor monthly.

Disclaimer

Model 1

Model 2 Revised

Model 3 - Revised

Still very much a work in progress. The turnover graph looks accurate now, but it's still difficult to determine if the strategy meets the monthly max turnover criterion. Maybe add a line showing mean monthly turnover as well?

It beats the SPY over this period, though you can't really tell from this tear sheet. Instead of using 'Specific Returns' (based on the Q Risk model), I'd suggest to use excess returns over the market, i.e SPY assuming that's the benchmark. Possibly volatility adjusted.

Vedran Rusman

Model #1 with latest NB

Arun

Model #7

Arun

Model #8

Arun

Model #9 (using optimizer)

Arun

Model #10

Arun

Model #11

Arun

Model #12

Model 2

Model 4

Leo M

Model1

Long Backtest of Model 4

Model 5

Model 3

Model 2

Model 4

Leo M

Model 2

Leo M

Model 3

Model 6!

Model 5

Arun

Model #13

Arun

Model #14

Arun

Model #15

Leo M

Model 4

Smart Beta 1

Smart Beta 2

Smart Beta 3

Leo M

Model 5

Model 6. Turnover still higher than required.

Model 3

Model 7

Youness AJ

Algo 1

Leo M

Model 6

Model 8

Baseline entry.

Algo 2

Algo 3

Leo M

Model 7

Model 9

Entry.

Model 4

Algo 4

Model 10

Model 5

Model 6

Model 7

Model 8

Leo M

Model 8

OlahasahalO

Smart Beta - Model 1

Hi!
Can I use a list of costraints to meet the challange's requirements?

order_optimal_portfolio(  
        objective=objective,  
        constraints=[  
            max_leverage,  
            dollar_neutral,  
            constrain_pos_size,  
            max_turnover,  
            factor_risk_constraints,  
        ]  
    )

How do I check that my factor is positive?

Something like

    screen = factor.top(300, mask=(qtu & ~factor.isnull() & factor.isfinite() & factor.ispositive()))

does not work

This way I can force the fact that I can go only long

Model 1.1

Vladimir Egorov

My first challenge - Model #1

Model 9

Model 1.2

Vladimir Egorov

My first challenge - Model #2

Juan Esteban Cepeda

Model 1

Model 1.3