Major mistake in the Capacity calculations

Continuation.

I spent a bit more time on this and I found a number of interesting cases.

1. Example $0 capacity when buying 10 SPY

I am simply buying 10 SPY at tick resolution, it is then claimed that this strategy has $0 capacity 

The next two backtests at seconds resolution show that 

self.MarketOrder("SPY",10)

and

for i in range(10):
  self.MarketOrder("SPY",1)

yield very different capacities. Observe how orders are placed at the same price 

and at the same time 

I do not expect returns, Sharpe ratio, and other statistics to be the same, however, capacity should not change because regardless of whether I buy 10 SPY once or 1 SPY ten times I end up with 10 SPY. 

Second backtest. I would have used tick resolution but that yields $0 capacity regardless.

ignore this. cannot delete post for some reason

Hi Pietro

As per my understanding, the capcity at that moment is caculated by the sum of 20% * Bar Volume / ( your order value / your whole account value ) for each order. So your bar resolution and order size do matter to your capacity. 

For the first algorithm, ADRU is indeed a very illiquid asset, only around 50k dollar volume on average for whole day, even less on seperate bars. There exists some time with very low volume when your trades executed with some illquid assets also. So the capacity is low for a reason.

For the second algorithm, maybe the single tick that your trade simulated is really like trade only 1 share, so the capacity is 0.

So my suggestion would be widening the resolution into minute or hour, excluding illquid equities by coarse universe selection or your manual selection process, using an execution model that is volume-oriented as well as check on spread.

Cheers
Louis

Also I missed one point, your order value = your order size * Ask price (long) or Bid price (short). So if the spread is extreme (like $0.1 vs $9999999), which is very likely for illiquid stocks as we always see market makers usually aren't placing their actual desire price on deck but fill up only when orders come in in actual cases, your order value will also become extreme and dragging the capacity.

Hi Louis, 

I am also confused by this metric. Am I right to assume that tick resolution has no volume, in which case, does mean that ALL strategies at tick resolution have $0 capacity? This seems like a strange metric to use.

Furthermore, can you explain the difference in the capacities 

self.MarketOrder("SPY",10)  #  $5.8 Billion

and

    for i in range(10):
      self.MarketOrder("SPY",1)  # $120 Million

20% * Bar Volume / ( your order value / your whole account value )

This formula suggests that buying 1 SPY shares 10 times would produce a higher capacity, however this not what I observe.

Best,

Asdas

Hi Asdas

1. The formula is 20% * filled price * Bar Volume / ( your order value / your whole account value )
    
2. He is using volume weighted average price execution mode, which has a restriction on filling volume, the remaining order unexecuted in that bar will leave to the next. That might caused the difference.
    
3. If that second has less than 10 share executed, the remaining ones also relay onwards, causing the difference
    
4. If for Resolution.Tick that tick has a big volume, everything is fine, but that only exist in ideal

Please if any staff could point out if this is valid.

Cheers
Louis

Hi Louis,

Thank you for your responses. Genuinely appreciated!

I am still confused. 

1. The fill price does not change (at least according to the orders tab).
2. I tried their code with self.SetExecution(VolumeWeightedAveragePriceExecutionModel()) removed and that made no difference to capacities. The difference is still $5.8 Billion and $120 Million.
3. All orders are executed at the same second (at least according the orders tab)
4. Okay I think that makes sense, I thought that SPY is as liquid as it gets but yeah

Help from staff would be very welcomed.

Best,

Asdas

Pietro Petitti,

The Major mistake is not in QC Capacity calculations but in your code which has 
Compounding Annual Return 0.769% and Drawdown 35.000%.
If you will use simple Buy and Hold SPY strategy you will get at different resolution
Estimated Strategy Capacity

Daily $1100000000.00
Hour $120000000.00
Minute $89000000.00
Second, $380000000.00
Second, VWAP $380000000.00
Tick, VWAP $380000000.00

Hi Vlad

I agree that lower resolution has lower capacity, however, I do not understand why buying 10 SPY once has higher capacity than buying 1 SPY ten times (at the same resolution), $5.2 Billion vs $120 Million. 

Fill price, bar volume stay the same (according to the info tab), and the only thing at changes is the order value. Can you explain this please? 

Best

Asdas 
 

Thanks for the feedback. There is a discount applied for “fast trading” to reduce the capacity of HFT strategies. This should only be applied when the direction “changes” (buy to sell) not repeated buy orders. We'll review and see if there's a good way to make it possible.

Hey Jared, Louis mentioned a formula to compute the capacity in a comment. Can you confirm it?

We'll write up a blog/documentation piece on it soon, but for now please answer these questions with a detailed understanding based on the underlying source code. I can't provide a simple Yes-No answer - capacity is more like an art than a science and we made an intelligent stab based on our best abilities and then open-sourced the work. 

CapacityEstimate.cs

SymbolCapacity.cs

I welcome informed debate and review of the source, without being informed is not constructive. We made these capacity estimates to help value alphas in the marketplace, and ensure investors were not diluted as alphas scaled. We'll continue improving the estimates over time.