InternFund Begins Live Trading

Hello everyone,

For the last month, Shile Wen and I, two Quantitative Developer Interns at QuantConnect, have been building trading strategies for a small in-house fund. We've spent the last several Friday afternoons working on this project and today we are excited to announce that we've deployed the fund's first two live trading strategies! In this post, we'd like to introduce ourselves, explain how users can emulate our workflow, and reveal the strategies we've deployed.

Introductions
Derek Melchin
I was first introduced to quantitative trading in the summer of 2018. Over the last two years, I've enjoyed spending my free time studying various quant finance topics and building personal projects. For the 2019-2020 school year, I accepted an executive position at the University of Lethbridge Finance Club. During my time there, I hosted several educational sessions and competed in the 2020 Rotman International Trading Competition as the algorithmic-focused team member. In the latter part of the academic year, I accepted an internship as a Quantitative Developer at QuantConnect. Now I utilize my background in software development and entrepreneurship to help the community and develop trading strategies.

Shile Wen
I was introduced to Quantitative Finance when I was browsing Quora, and came upon a post about the Medallion Fund, the legendary fund ran by Jim Simons and RenTec. I thought it was fascinating, and that’s what got me hooked. I currently attend the University of Washington, with a major in Computer Science. At UW, I partake in the Algorithmic Trading Club at my school. I hope to help the community with their issues and develop trading strategies (we have a Strategy Library for those that want to view strategies developed by interns) for the community.

Team Workflow
Shile and I both work in different countries, yet we've been able to remotely manage the InternFund together via the project collaboration feature. In short, we performed our research independently but merged our strategies together into one project after we validated their individual performance. The Algorithm Lab made deploying these strategies very simple. For an overview of the point-and-click process, check out this quick video.

InternFund Strategies
A concern some users have, and rightfully so, is that our staff may take their private strategies. Rest assured, we don't. Our terms of service and privacy policy explain this at great lengths. Shile and I don't even have access to users' private projects or the Alpha Streams algorithms. The strategies we researched and deployed for the InternFund can be found in public research papers.

The strategies we deployed are:

Algorithm
We merged the two strategies into one classic-style algorithm. A backtest of our live deployment is attached.

Future Updates
We hope to keep the community updated with the performance of the InternFund throughout the rest of the summer, so stay tuned!

Best,
Derek Melchin

The material on this website is provided for informational purposes only and does not constitute an offer to sell, a solicitation to buy, or a recommendation or endorsement for any security or strategy, nor does it constitute an offer to provide investment advisory services by QuantConnect. In addition, the material offers no opinion with respect to the suitability of any security or specific investment. QuantConnect makes no guarantees as to the accuracy or completeness of the views expressed in the website. The views are subject to change, and may have become unreliable for various reasons, including changes in market conditions or economic circumstances. All investments involve risk, including loss of principal. You should consult with an investment professional before making any investment decisions.

var scheduledTime = _algorithm.Time.AddMinutes(1);
_algorithm.Schedule.On(
_algorithm.DateRules.On(scheduledTime.Date),
_algorithm.TimeRules.At(scheduledTime.TimeOfDay),
() =>
{
StopLossTicket = _algorithm.StopMarketOrder(_security.Symbol, -Quantity, StopLossPrice, tag: "ATR Stop");
_algorithm.CheckToCancelRemainingEntries();
});

Derek Melchin INVESTOR

QuantConnect | July 2020

Day 1: On the Up

Upvote

5 people upvoted this

Pi..R INVESTOR

July 2020

Very cool internship project guys, thanks for sharing :)

2 people upvoted this

Jovad Uribe INVESTOR

This is amazing!

Jon Quant INVESTOR

Nice! Good luck with guys.

Week One: A Good Start

Hi everyone,

The live deployment has been profitable throughout its first week!

We researched some new strategies today but not find anything worth adding this week. Instead, we've decided to increase the allocation ratios of the two current strategies. This adjustment raises the Sharpe ratio of the backtest to 1.95 while maintaining a max drawdown <$1K. The <$1K drawdown is vital as the InternFund is only allowed a $1K drawdown.

Going forward, we are excited to research more strategies to add to our portfolio. We have a few in the pipeline, but are always looking for more. If our readers have anything they would like us to test, feel free to add suggestions to this forum thread.

See the attached backtest for the latest deployment.

4 people upvoted this

HughStryker INVESTOR

Thank you for sharing, I look forward to following this thread.

I'd be very interested to see a QC implementation of Chris Cole's Dragon portfolio, i.e. 20% allocation to equities, bonds, trending commodities, gold and active long volatility. It was created in response to the typical 60/40 split so it would be nice to see how the two strategies compare.

Shile Wen INVESTOR

UW CFRM | July 2020

Hi Hugh,

Thank you for the suggestion, we'll keep it in mind.

Best,
Shile Wen

1 person upvoted this

Ernest Shaggleford INVESTOR

Hi Derek and Shile,
thanks for revealing your strategies that you are trading live.
On the weights used for the 60:40 algo set of tickers:

self.weight_by_ticker = {'SPY': 0.6, 'AGG': 0.4, 'VXX': 0.1}

The total weight for this set is 1.1 which is then scaled by the allocation ratio of 0.5 so it would be 55%.
But as the TLT (treasury proxy) algo allocation ratio is 50% then the total portfolio allocation with both algos could potentially be 55% + 50% = 105% during the Treasury holding period, which doesn't seem correct.
Have I missed something?

Thanks,
ES.

Week 2: The Hunt for the Next Strategy Addition

Hi Everyone,

This week, we’ve been researching new strategies for the InternFund. One strategy we researched is a Gap Down strategy, where we buy stocks during the day that surpass the previous day’s close if they open down 1% from the previous day’s close. However, the backtest (attached below) was not promising, so it will not make it into the InternFund. We also developed an option spread strategy, however, we deemed that the drawdown was too large, thus, this strategy will not be added either.

Here is the performance of our portfolio for this week.

Ernest Shaggleford pointed out that with our current ratios, we’d have 105% allocation during periods when we hold TLT. We confirmed that our account has access to overnight margin, which means going over 100% allocation will not affect our algorithm.

As for next week, we will implement HughStryker's suggestion and keep researching new strategies.

Hi Shile,
re: "We confirmed that our account has access to overnight margin, which means going over 100% allocation will not affect our algorithm":
Out of curiousity, won't the cost of margin have some impact on the P&L?

Thanks,
ES.

Thanks for the update, Shile!

Here is another strategy idea, I'm too much of a beginner to accurately describe it, but I'd say it's a L/S portfolio using institutional ownership as a factor. There's a variation which uses long/short interest instead.

BofA version
ZeroHedge variation

Just a quick update: InternFund closed at a new high today!

Stay tuned for a more comprehensive update this Friday.

3 people upvoted this

Week 3: Researching More Strategies

Hi Everyone!

This week, we researched more strategies for the InternFund. One strategy we tested this week was HughStryker’s suggestion, Chris Cole’s Dragon Portfolio. Although the drawdown is relatively low compared with other strategies, it is still too high for us, so we decided not to add it. An implementation can be seen in the attached backtest.

Derek implemented a Trailing Breakout Strategy with TSLA. It performs well throughout 2020. However, extending the backtest start date further into the past decreases the performance. For the InternFund we are specifically looking for strategies with a long track record of consistent returns.

As for the InternFund performance, we are at a new equity high! Five days ago, we entered our first TLT trade, and we have exited TLT today, for a gross profit of $76.54. Here is the graph of our equity:

Ernest Shaggleford mentioned that margin will affect our P/L, which is true, however, we have very low margin interest, and we only hold TLT for five days out of the month, so the costs are negligible. HughStryker offered another strategy of incorporating institutional ownership. Unfortunately, we do not have access to this data.

Thanks to everyone for tuning in to our weekly update, and special thanks to HughStryker for the Dragon Portfolio strategy suggestion! Stay tuned for more updates, and feel free to suggest more strategies.

Josh M INVESTOR

August 2020

Interesting; cool discussion and glad to see it's working live!

Tamim Fund INVESTOR

Shile,

The Gap Down strategy seems to be executing at "stale prices," resulting in there being no price gain or loss other than the commission, since you are comparing the current days close to the previous days close and placing a buy order after market closes, instead of comparing intraday minute or hourly price data. Could you look at that?

Thanks for the update!

QuantConnect | August 2020

Hi Ernest and Tamim Fund,

We appreciate the constructive feedback. However, the InternFund is limited to only trading equities. Thus, we cannot implement the suggestions #2 & #3 from Ernest above. As for the Dragon portfolio changes, we will take a closer look at this today.

In regards to the stale prices in the Gap Down Strategy, we will work to resolve this today as well.

Week 4: Up To Three Strategies!

We are excited to announce that the InternFund generated profits for another week. Here's the latest screenshot of our live track record.

Today, we spent some time fixing the "stale price" issue with the Gap Down strategy. The issue surfaced because the universe resolution was set to daily in the original algorithm while the strategy was trading off intraday prices. Here is the backtest of the strategy after fixing the bug. Thank you Tamim Fund for reporting this issue.

This week, we also spent some time implementing the Turnaround Tuesday strategy. As can be seen from this backtest, the strategy has performed quite well over the last few years. Allocation 33% of the portfolio to the strategy generates a 58% win rate, >1 Sharpe, and a mere 1.6% drawdown. We are confident enough in these results that we've decided to deploy this strategy to live-trading within the InternFund algorithm. To incorporate the strategy, we reduced the allocation ratio of the 60/40 strategy to 30%, allowing the Turnaround Tuesday strategy to have a 20% portfolio allocation.

We found out today that the post of our latest live deployment above actually quits early because it reaches our max drawdown limit of $1000. After increasing the max drawdown limit so the backtest completes, the original max drawdown was 6.3%. With the Turnaround Tuesday strategy incorporated into the algorithm, the new combined strategy experiences a 3.8% drawdown. The Sharpe ratio is also increased from 1.64 to 1.93. We will of course only deploy the new strategy with the $1000 max drawdown limit, but a complete backtest of the InternFund algorithm (with an extended drawdown limit) is attached below.

For those readers who may have missed the discussion thread introducing our newest team member, we'd like to introduce Jovad to this thread. He implemented an interesting strategy this week while researching for the InternFund. It's a Fibonacci Option Straddle strategy that sets a retracement at 30% below the 5 day maximum of the underlying security. Every two hours a scheduled event is triggered checking if the nearest expiring and strike price option's bid price has met or crossed the set retracement. If this is true, then we enter a long position. If it's false, we liquidate our position, or we do nothing. Lastly, the universe was selected based on equities that have had a historical Implied Volatility <=35% for the options with the nearest strike price. The strategy performed well but it would cost the fund $400/month for an options data feed in our account and would cancel out returns.

Thank you everyone for following along with our live track record. The support and constructive criticism gets us excited to work on this each Friday. As always, we encourage our readers to stay involved by requesting trading strategies for us to test for the InternFund.

Stay tuned,
Derek Melchin

For the option straddle strategy, it only gets the option with the absolute lowest strike price. The expiry doesn't matter since it is sorted by the strike price afterwards.

Could you explain the logic behind this strategy? It seems that it relies on the variety of the underlying options, since only very very deep ITM calls will be worth between .7 and 1.3x the security price, instead of relying on some trading principle.

I just ran it, and the reason it works is because 99% of the time, the only call options with a low enough strike price to fit the parameters is apple's and sometimes, GS, V, TGT, and MSFT. Apple has been going up for the past decade. About 80% of the orders over the past 5 years is just Apple. In the past year, there has been more Microsoft, but it is still mostly Apple.

If I am wrong, please correct me.

Lars Klawitter INVESTOR

January 2025

To Yuri's earlier point of there being a difference between backtests and live trading as to when the stop orders are placed:

I've been running the strategy (using a larger universe and shorter opening range as per my previous post) on IBKR paper trading (i.e. on an actual IBKR account, not via QC paper trading) and as Yuri suggested, the stop orders are at times placed within the same minute. Mostly 20-40 seconds after the entry:

in one case the stop order was placed the same instance as the entry:

I tried second resolution, but that seems impractical given the large universe size.
So this is my attempt at an artificially delayed stop order placement:


                var scheduledTime = _algorithm.Time.AddMinutes(1);
                _algorithm.Schedule.On(
                    _algorithm.DateRules.On(scheduledTime.Date),
                    _algorithm.TimeRules.At(scheduledTime.TimeOfDay),
                    () =>
                    {
                        StopLossTicket = _algorithm.StopMarketOrder(_security.Symbol, -Quantity, StopLossPrice, tag: "ATR Stop");
                        _algorithm.CheckToCancelRemainingEntries();
                    });

I'm not a C# coder, so I definitely don't know what I'm doing. Backtests with this code produce by and large comparable results with the original code, so I'll try paper trading next.
Would the above code change make sense to you?

Yuri Lopukhov

Yuri Lopukhov | January 2025

I'm not a C# coder myself, I suppose your solution should work, thank you for sharing. Yes, I've also noticed significant increase of returns when reducing entry to first minute (and also increasing actual leverage used). But it also increases chances of this issue affecting backtests, so not really sure if we can trust those results yet.

Actually, your solution will delay stop loss order by exactly one minute always, while on minute resolution stop orders would be delayed until the next minute start, i.e. if entry was filled at 9:36:20, your solution will place order at 9:37:20, while backtest would do that at 9:37:00. Maybe you could replace seconds with zeros for scheduledTime variable like this: scheduledTime = new DateTime(scheduledTime.Year, scheduledTime.Month, scheduledTime.Day, scheduledTime.Hour, scheduledTime.Minute, 0);

I don't know about you, but for me this change didn't improve things at all )

Lars Klawitter

Lars Klawitter | January 2025

I didn't actually expect the change to improve performance in backtesting, but it is there to avoid the stop losses to be placed too early in live trading, a phenomenon that we would not see in backtesting. So the artificial delay not changing the performance in backtesting would rather give me confidence in the viability. Wasn't that the concern you raised originally, when you pointed out the discrepancy between backtesting and live trading order placement?

and where the very short opening range is concerned: I guess that's the biggest concern. I've been paper trading on an IBKR paper trading account for three weeks and so far it behaves very similar to a corresponding backtest, but the real acid test will be actual live trading (which only makes sense for cash books >$100k on brokerages like IBKR with trading fees)

Yuri Lopukhov INVESTOR

So far I can see two issues:

VolumeSMA is not warmed up when new symbols are added to universe, which means these symbols will not be selected for the first 14 trading days in universe
On Second resolution it chooses different symbols to scan and trade. Perhaps this is related to the first issue.

I can't fix the first issue in C#, so I guess I will switch to Python version unless somebody else fixes the C# version and share it. Not sure if fixing it will improve results as well.

Derek Melchin

QuantConnect | January 2025

Yes, we opted to not warm up the VolumeSMA to minimize the number of code lines for readers.

... and to increase the execution speed.

Jake Chen

Jake Chen | January 2025

I've mentally accepted issue 1... My core assumption is that the top 1000 symbols sorted by volume will not change drastically from month to month (when symbols are added/removed) and thus the number of these lagged-14 symbols shouldn't be too much. I suppose a way to get around this would be to simply increase from 1000 to like 1250 or something, and mentally write-off the bottom 250 as constantly rotating and never in-play.

Although I don't have data to back up that core assumption haha, so I'm be happy to be challenged there.

Jack Pizza INVESTOR

Sum Sum | January 2025

define: “does really well with 1 minute” ….. 2000-2002 still utterly collapses fail to see well really well fits in….

or does really well when overfitting super hard?

hmm, I hadn't tested timeframes as early as 2000 before and you're right, the strategy doesn't perform as well as in later timeframes (2010+) and the whole period of July 2001- end of 2002 is basically flat, but I'm still getting a Sharpe ratio >1 over the whole period and only 10% drawdown. But yeah, I still distrust the results, I guess I've got quite a bit of overfitting going on and I'm uncertain as to how the short opening range will do in live trading.

Quant Stratege INVESTOR

These backtests are not representative of live performance. When adding slippage, it can be significant at the open due to volatility, small-cap stocks, and using stop orders, making the results much less appealing.

Just add this line when adding securities:

security.set_slippage_model(MarketImpactSlippageModel(self))

You're right. I had previously only simulated very small constant slippages, but MarketImpactSlippage has quite a savage effect…
If it looks too good to be true…

This slippage model is for strategies with large market impact. If you are planning to trade hundreds of millions, sure, this strategy might not work for you. For smaller accounts, market impact on top US stocks will be much smaller, especially if universe_size is smaller than 2000. Also, that requires proper parameters, defined here: https://www.quantconnect.com/docs/v2/writing-algorithms/reality-modeling/slippage/supported-models#05-Market-Impact-Model If you just use default parameters, you probably just assume much worse default scenario.

looking at the capacity, it is somewhere between $100k and $500k. If you add to that the fact that with IBKR's fee structure, anything below $100k is not viable either, you end up with a fairly narrow band of viable cash book sizes - which might not be a bad thing!

the capacity can be slightly increased by filtering for minimum dollar values either in the universe selection and/or when selecting the "stocks in play". That only goes so far before eroding all alpha. I've been trying a dynamic filter using a multiple of the portfolio value.

AleNoc INVESTOR

Alessandro Nocentini | January 2025

I noticed a difference between the backtest and the selection with live data (data provider QuantConnect). What data provider do you use for live trading?

Jake Chen | February 2025

I'm seeing the same difference. I use TradeStation. Were you able to resolve this issue?

Evan lightner INVESTOR

I don't know where to start, I'll have to write a whole post and how I'm planning to implement this strategy. But first of all thank you @derek for sharing this (and of course for everything you've done with QC -its been a game changer for me).

But first of all, I just want to comment with one simple non-coding question….

Why are a good amount of people HATING on this strategy? I understand the backtest cherry pick , but for a bare bones boiler plate ( i messed around with some variables like holdings and percent risk and still got good results) , this is the best algo I've seen in a while, especially for being shared so freely - not to mention brand new piece of research in the community.

Is there just some sort of deep skepticism in the quant community at large I guess, inherently? I suppose that fits!

Kevin Patterson INVESTOR

February 2025

Does anyone else end up with very different symbols when running the exact same algo in backtesting vs live (IBKR paper trading)? For example when it is set to choose 20 symbols, I'll see maybe only half overlap –- and of course the ones going positive for the day in the backtest are the ones not picked up live 😅. Seems the relative volume calculations aren't exactly the same and it doesn't take much to move the symbols you get. This is my first foray in to large universe algo's, is this type discrepancy common with large universe backtests or is there settings to help make it line up better with live? S

Some folks were asking about sized nodes: after trying the python version on IBKR (with a few mods) an L1-2 node will make it through one day (it crashes after close though, so likely you need the next level up if you dont want to restart daily)

Thanks for all the python related posts, even if I don't end up trading it, the algo has been super helpful for learning more about the QC code and had some good recipes in it that I think would be helpful for any algorithm.

Yuri Lopukhov | February 2025

Note that the universe selection is happening on the first day of the month in backtesting. While on live it will run upon deployment. Also, indicators with recursive formulas, like ATR may have different values on different periods even when these periods are larger than their "period" (because it is not actually a period, but a factor). That are few things that come to my mind, I am still running it on paper and didn't dig into this yet.

Evan lightner

evan lightner | March 2025

I came back to this thread to ask this literal exact same question! I'm trying to understand, *why* are the symbols different in the IBKR paper trading algo versus the QC paper brokerage (back/forward testing)? Are you saying if I launch both the QC paper brokerage live node (or backtest) *and* the IBKR (paper or cash) live node at the exact same time at the beginning of the month, they will *then* both trade the same symbols? Additionally, I've been thinking (before seeing this comment) that the VolumeSMA not being warmed up when new symbols are added (as you confirmed above) was the reason for this discrepancy. But is that not relevant?. What is the reason VolSMA this isn't being warmed up anyway (and therefore needs 14 days)? Is there not a way to do this, or was that just a choice for readability/speed? Thanks Yuri (and Kevin)..

Yuri Lopukhov | March 2025

The reason VolumeSMA is not warmed up is because it requires 14 days of minute data for all securities in the universe for initial warm up, and then for each security added to universe (but that is a much smaller issue). This greatly increases start up time and ram requirements, which may cause time out delay on live instances and perhaps even in backtesting. Not regarding to this issue specifically, I should mention that backtesting on minute resolution provides very different results from second resolution even after all technical issues are resolved, because this strategy trades on very volatile market opening and order fills are very different on minute and second resolutions.

Platform

Radically Open-Source Algorithmic Trading Engine

Join Our Discord Channel

Quarterly Open-Source Trading Competition

Draft Discussions

Bookmarked Discussions

SEARCH DISCUSSIONS

369,100 Quants.

VOTE FOR UPCOMING FEATURES

JOIN OUR Community MAILING LIST

IN THIS RESEARCH

PARTICIPANTS

Actions

Join QuantConnect for Free