Forums  > Pricing & Modelling  > Where to find inspiration for signal ideas  
     
Page 1 of 1
Display using:  

TheNikster


Total Posts: 3
Joined: Jun 2020
 
Posted: 2020-06-14 20:39
I get the order queues at top 5 price levels on ask / bid side for a futures contract every ~.1ms, and I need to predict the return on a ~1 minute time horizon. The setting here is online, so i'm not able to train my model on a large amount of existing data, I only receive updates one by one.

Are there any places which have signal ideas that I can get inspiration from? I already have the basics, like book pressure / mean reversion and am getting 9% R^2 in my model, but I'm hitting a roadblock here.

bookpressure
Banned

Total Posts: 3
Joined: Jun 2020
 
Posted: 2020-06-19 21:09
one of Best place to get inspiration is Changshu New & Hi-tech Industrial Development Zone. We help many American company like CITADEL INVESTMENT GROUP outsource Signal Ideas like book pressure / mean reversion.

Some of Our Achievements include :

- only ISO 9001:2005 certification in Yangtze Delta core region of Jiangsu
- sustainable zero emission industrial production
- we build volatility model condition on large price move for top 5 index option trading company
- we invent use malform message to cheat matching engine to mass cancel order very fast
- We Use special model class call RANDOM DECISION JUNGLE to achieve two-class prediction for Double Uptick about 87.1%
- We build NYSE D QUOTE model that let big firm BUILD BIG IMBALANCE at closing that not possible !!

Our collaboration process to increase R^2 for inspiration include watching videos on xvideo.com

bookpressure
Banned

Total Posts: 3
Joined: Jun 2020
 
Posted: 2020-06-19 21:09
one of Best place to get inspiration is Changshu New & Hi-tech Industrial Development Zone. We help many American company like CITADEL INVESTMENT GROUP outsource Signal Ideas like book pressure / mean reversion.

Some of Our Achievements include :

- only ISO 9001:2005 certification in Yangtze Delta core region of Jiangsu
- sustainable zero emission industrial production
- we build volatility model condition on large price move for top 5 index option trading company
- we invent use malform message to cheat matching engine to mass cancel order very fast
- We Use special model class call RANDOM DECISION JUNGLE to achieve two-class prediction for Double Uptick about 87.1%
- We build NYSE D QUOTE model that let big firm BUILD BIG IMBALANCE at closing that not possible !!

Our collaboration process to increase R^2 for inspiration include watching videos on xvideo.com

bookpressure
Banned

Total Posts: 3
Joined: Jun 2020
 
Posted: 2020-06-19 21:09
one of Best place to get inspiration is Changshu New & Hi-tech Industrial Development Zone. We help many American company like CITADEL INVESTMENT GROUP outsource Signal Ideas like book pressure / mean reversion.

Some of Our Achievements include :

- only ISO 9001:2005 certification in Yangtze Delta core region of Jiangsu
- sustainable zero emission industrial production
- we build volatility model condition on large price move for top 5 index option trading company
- we invent use malform message to cheat matching engine to mass cancel order very fast
- We Use special model class call RANDOM DECISION JUNGLE to achieve two-class prediction for Double Uptick about 87.1%
- We build NYSE D QUOTE model that let big firm BUILD BIG IMBALANCE at closing that not possible !!

Our collaboration process to increase R^2 for inspiration include watching videos on xvideo.com

nikol


Total Posts: 1124
Joined: Jun 2005
 
Posted: 2020-06-22 13:55
Money losses are the best inspiration and motivator to find new signals.

Azx


Total Posts: 43
Joined: Sep 2009
 
Posted: 2020-06-22 15:32
1 minute return of what? Mid-price?

TheNikster


Total Posts: 3
Joined: Jun 2020
 
Posted: 2020-06-25 02:44
Yes

EspressoLover


Total Posts: 432
Joined: Jan 2015
 
Posted: 2020-06-25 16:21
> Our collaboration process to increase R^2 for inspiration include watching videos on xvideo.com

Ahh... So that explains the web browsing habits of a former co-worker of mine...

Anyway, regarding OP's question, I'd take a step back. Who are the ultimate consumers of the signal? There's going to be a big distinction based on that. A market making system will potentially want to use very different alphas than a directional liquidity taking strategy or an execution algo for for large positions.

It sounds like you're fitting alphas on a dataset of arbitrary time slices. Which is fundamentally a rebalance-based approach. At that horizon, it's usually better to think in terms of discrete events that trigger trading actions. You can identify candidate sets that constitute potential events, and specifically fit your signals at those points. Rather than random points in time. If you're providing liquidity, then those events are obviously times when liquidity gets filled. But even if you're taking, there's probably a narrow range of events that constitute the vast majority of your triggers. Things like a narrowing of the spread, new level formation, large trade at the touch or a tick in the index futures or leading instrument.

Also, what makes you so confident that the limitation is the lack of signal? 9% out-sample R^2 on 1-minute returns seems pretty decent to me. (Depending on liquidity, microstructure, tick size, book thickness, volatility, etc.) What would you be happy with? Do you have benchmarks against competitors? Are you sure this is the bottleneck? Are you sure there isn't lower-hanging fruit by reducing latency, or smarter order placement, or better risk management, or lowering trading costs, or a wider instrument universe?

Now to actually answer your question... It seems like the biggest thing missing from your model is accounting for price discovery that happens in other instruments. Say, you're trading NQ, then you definitely want to be incorporating the action in ES. And most likely the bond futures, VX, and even the cash equities market if possible. You can start by looking up the academic literature on the Epps effect for to help get started.

You mention book pressure, but that means very different things to different people. At the most basic level, you're just taking the difference between the mid-price and weighted mid-price. And that works great, but you can definitely get a lot more sophisticated. If you have order book data, you can qualify the individual orders that make up the queue. Depending on the instrument, the deep book may have predictive value.

And most importantly the historical evolution of the book is as, if not more important, than the instantaneous state of the book. For example a market's that quoting 500x500 then has a market buy order for 400 lots is a lot more bullish at that instant then a market that's been quoting 500x100 for the past ten thousand milliseconds. This kind of segues into a third source of signal in that you can profile the order flow to derive signal. I'd check out some of the archives at the blog Mechanical Markets for some more ideas.

Good questions outrank easy answers. -Paul Samuelson
Previous Thread :: Next Thread 
Page 1 of 1