Iain Dunning
@iaindunning
Followers
2,653
Following
356
Media
84
Statuses
547
Head of AI @ HRT ( @wehrtyou )
New York, NY
Joined January 2013
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Michigan
• 312235 Tweets
Arsenal
• 267092 Tweets
Miami
• 120473 Tweets
Tennessee
• 103590 Tweets
#StarAcademy
• 81530 Tweets
Bournemouth
• 81064 Tweets
Nico
• 77614 Tweets
Francis
• 53486 Tweets
Usher
• 53380 Tweets
Celta
• 45023 Tweets
Lizzo
• 42801 Tweets
Norris
• 39908 Tweets
Verstappen
• 34920 Tweets
Romero
• 32483 Tweets
Russell
• 31030 Tweets
Nebraska
• 29568 Tweets
Hamilton
• 28582 Tweets
Atlético
• 27577 Tweets
Alonso
• 27397 Tweets
Modric
• 26005 Tweets
GARIME X LA OPORTUNIDAD
• 19181 Tweets
Bama
• 17979 Tweets
#PFLSuperFights
• 16339 Tweets
Ancelotti
• 15475 Tweets
Renato
• 13817 Tweets
Hulk
• 13811 Tweets
Tigre
• 12351 Tweets
Arnold Palmer
• 11206 Tweets
Please apply! I am especially interested in PhDs this year - don't put off applying until the end of the year (speaking from my personal experience...)
Summer 2025 internships and full-time roles are live! 🎉 From Algorithm Development to Software Engineering to internships in NYC and Singapore, work on challenging and impactful projects. Visit to see all our new roles.
1
8
89
10
13
216
My AI team is always on the lookout for excellent people who want to build cool things. Plenty of data, plenty of compute. Markets are a challenging domain where off-the-shelf approaches don't work or are at least very suboptimal, which adds to the fun. DM me, email me.
17
17
202
Speaking as a former DM engineer, you should disregard this. DM (and OAI) was built by talented engineers, many with PhDs. This person has been at DM 2 months, & the real red flag is having time to tweet multiple shallow takes a day. Don't let people gate keep you from AI!
6
3
136
Not sure if "extremely good" is necessary but, in my experience at DM and now at HRT, one can only contribute at the highest level if you can implement your good ideas effectively (your own time) and efficiently (machine's time).
coming up with good ml research ideas is significantly easier than implementing them in a complex codebase — the returns to being an extremely good engineer are super high
29
85
1K
0
1
44
Will be attending MIT Career Fair next week, come say hi! First time in 5 years, should be fun.
1
0
30
2
2
26
Best thing I did in my PhD!
Iain Dunning (
@iaindunning
), Joey Huchette, and Miles Lubin
for the paper
JuMP: A Modeling Language for Mathematical Optimization ()
@JuMPjl
2
12
62
1
0
29
Come say hi in San Jose next month!
Hear Hudson River Trading (
@wehrtyou
) speak about how they use
#AI
for price discovery and trading in
#quantfinance
. Join the session at
#GTC24
on March 19.
7
6
24
0
1
25
Google just open-sourced their LP solver Glop, added SCIP-powered MILP solver to Google Sheets, + a solver API
#orms
http://t.co/B1fMwQgDDd
2
27
27
Best paper I've read in a while, and captures why a lot of the optimisation literature is pretty damn terrible
0
7
24
Excited to meet people at ICML, we know there are a lot of really interesting opportunities right now for talented AI researchers and engineers so please come by and let me pitch you on why you should consider working with us.
Innovative strategies are essential to harness the complex data of financial markets. For
@iaindunning
, head of HRT AI Labs, it's the perfect playground for advanced
#ML
. From July 22-24, stop by booth 311 at
@icmlconf
in Vienna to learn more and connect with our team.
#ICML2024
5
1
27
2
1
21
The tweets this is responding to, and replies therein, reveals how overloaded "quant" is, causing much confusion. Add in intentional opacity around the capacity of various strategies, and a common misunderstanding that all "smart" participants are seeking max total return
My dude what have you not understood from what I’ve just said.
I spend 8-10 hours/day building Sharpe 10 HFT strategies. I have zero desire (and certainly lack of tech and other obstacles) to do so after that in my free time.
Also my employer would have a look at that if I did.
17
6
287
1
1
19
Watching
@MIT
Prof Steven Johnson's talk about how
@JuliaLanguage
&
#python
can be friends (2 lang prob!)
@EuroSciPy
0
6
17
Talk "Fast & Flexible Linear Algebra in
@JuliaLanguage
" by
@ANoackJensen
@ Stanford ICME -
@ProjectJupyter
powered!
1
5
16
At this week's
@wehrtyou
AI Labs (HAIL)
@bnzyx1
(who just started this fall) presented "Real World Games Look Like Spinning Tops" by
@wojczarnecki
@maxjaderberg
et al - great visualizations, and fun to ponder how complex systems like financial markets and can be analyzed this way
0
5
14
Exciting day!
JuMP 1.0 is out! Use JuMP to route school busses, schedule trains, plan power grid expansion, or optimize organic milk production! (Just examples of successful applications.)
5
73
350
0
0
14
New: GraphLayout.jl force-based graph layout algos + SVG plotting w/ Compose.jl
#JuliaLang
http://t.co/DKCY41Q3xj
0
5
13
Had honor of presenting some of the AI research I've been doing at
@DeepMindAI
and
@wehrtyou
at
@WestPoint_USMA
this week. The cadets were a great audience with perceptive questions about AI now and in the future, incl. safety, economic impact, and how to learn more about DL+AI.
0
0
13
Excited to be at
#NeurIPS2018
! Will be talking briefly at the Expo on Sunday on behalf of
@wehrtyou
about some issues and opportunities around applying
#DeepRL
to trading. 3pm, room 517D.
0
1
13
Last week Nick Watters (incoming
@mitbrainandcog
PhD, ex-
@DeepMindAI
) dropped by HRT
@wehrtyou
to give a presentation about unsupervised representation learning at our bi-weekly paper reading group - and challenge our
#foosball
pros too!
0
1
12
Haven't had much opportunity to
#JuliaLang
for a few years, but still fun to tinker - just updated to 0.3.0, once I figured out how package registrations work these days... Just a few years ago I was running the package listing website!
0
1
12
Impressed with
#Numba
performance, simulation speed up by 30x + moved bottleneck back towards
@TensorFlo
where it should be!
#python
#ai
0
1
12
(and yes I invest primarily in low cost index funds)
That not all "smart"/"quant" are optimizing for the exact same objective is one of the things that makes markets extra interesting to me from an AI perspective. E.g. ideas like self-play (AlphaGo) don't really map well.
0
1
10
At this week's
@wehrtyou
AI Labs (HAIL) Paper Survey Club
@marckkhoury
presented "Switch Transformers" by
@LiamFedus
@barret_zoph
@NShazeer
. Offers a way to embiggen models without blowing out computational costs - trillion parameter models!
0
1
11
"...one effective approach to addressing climate change is contributing to the development of
@JuliaLanguage
"
0
17
11
#ORMS
magic =
@JuliaLanguage
+
#JuliaOpt
JuMP +
@ProjectJupyter
IJulia +
@dcjones
Gadfly +
@g0wda
Interact.jl = paper
http://t.co/5sy1IWyVHn
2
5
10
@lemergenz
Going from state of "not doctor" to "doctor" definitely is arbitrary/social construct, but the work the average person does to get that label is quite real.
1
0
11
New paper about
#JuliaOpt
JuMP & how w/
@JuliaLanguage
it enables exciting new ways to explore optimization models
http://t.co/5sy1IWyVHn
0
8
10
At this week's
@wehrtyou
AI Labs (HAIL) Paper Survey Club I presented "Smart 'Predict, then Optimize'" by
@Adam235711
@paulgrigas
- SPO very relevant to trading, where putting effort into what you predict (loss/label) is just as important as how you predict it (linear/NN/RF/etc)
0
0
10
Strong name, fun idea
The gradient is a locally greedy direction. Where do you get if you follow the eigenvectors of the Hessian instead? Our new paper, “Ridge Rider” (), explores how to do this and what happens in a variety of (toy) problems (if you dare to do so),.. Thread 1/N
4
71
583
1
0
10
Read about some of the fun stuff my interns did this year, and apply to be one next year!
0
0
9
Slides for my talk
@BayAreaJulia
re
@JuliaLanguage
's optimization packages
#JuliaOpt
#orms
http://t.co/11DarxT36T
2
4
8
.
@wehrtyou
is hosting a "Women with HRT" Tech Summit on 5/21 & 5/22! Our very first virtual summit for women/nonbinary undergraduate students. Expect talks (incl. about AI!) workshops, networking, social events (+ swag naturally) See for the details/reg.
0
3
7
Something to watch,
@beenwrekt
's posts on RL/"policy gradient" still excellent reading I often recommend.
This semester, I’m back to live blogging my course lectures. I’m teaching Convex Optimization in the Age of LLMs.
7
75
648
0
0
7
@YasminMoslem
I found contributing to open source projects the most impactful thing. Not only did it help improve my engineering skills, it also creates a portfolio of work beyond papers which can help in getting your first job.
1
1
8
Learning about how
@DidiChuxing
apply
#RL
to matching problems at
#NeurIPS2018
Expo. Hudson River Trading
@wehrtyou
will be up next at 3pm in room 517D!
0
1
8
I enjoyed this from
@macrocephalopod
, although having talked to a lot of undergrads on campus and through internships I do think there is a good chance this person is "gmi". This tweet is even more interesting for the replies though, many have big ngmi energy...
2
0
7
Had a lot of fun teaching a class on NLP techniques, old, new, and state-of-the-art at
@wehrtyou
this week. Interesting to reflect on the improvements the last decade brought, hopefully another exciting decade ahead!
0
0
8
New post: Naval Warfare with JuMP
#JuliaLang
http://t.co/KiJFtccqvt
- use
#orms
to lead your forces to victory - or plot some circles&lines!
0
5
7
My Capture-the-Flag work with
@maxjaderberg
@wojczarnecki
et al from
@DeepMindAI
has been published today in
@sciencemagazine
! Updated blog post with great results (and visualizations!) on real Quake 3 maps + playing with humans
0
1
7
@JonathanUesato
One great thing about applied research in industry is there is no tension - the only important thing is to make progress, and there are no prizes for overcomplicating your method/doing something novel for sake of it
0
1
7
This morning at 10:30! I'll be doing the intro. We also have a booth, please drop by for limited edition HRT+NeurIPS t-shirts.
Join HRT at
@NeurIPSConf
(Booth 518) in NOLA! Check out our talk, “Machine Learning and Optimization for Automated Trading at HRT” on Monday Nov 28 in Theater B. Learn about the tools we use to make trading decisions, from Deep Learning to Bayesian Optimization. 🤖
#NeurIPS2022
0
0
3
0
0
7
That cold sweat moment when realize you've been plotting with wrong x-axis all week & are scared to see actual results
#justacademicthings
1
0
7
I heard multiple versions of this error at
#NeurIPS2022
, e.g. from a student who "made a 3 layer neural network that got 60% accuracy predicting price changes", but with the error below. Good reminder, quant finance is a systems problem, data to model to execution - makes it fun!
0
0
6
Shower thought: high-profile multiagent AI research is focused on zero-sum games (SC,Dota,Go) - killing your enemy!
IRL, interactions where both parties "win" are normal. E.g., financial markets: due to time/liquidity prefs, most(?) transactions are +ve utility for both parties
1
0
7
@typedfemale
Glad you enjoyed! PyCon attendees enjoyed it too
@chrisjrn
Oh, we get it. We have 492 of these bags in stock. DM us a receipt for a $5 minimum donation to
@CallenLorde
and we'll send you one.
#ProudAtHRT
🏳️⚧️
8
30
133
1
0
6
At today's
@wehrtyou
Paper Survey Club, we're going to be looking at "MetNet: A Neural Weather Model for Precipitation Forecasting" by Sønderby,
@NalKalchbrenner
, et al. . A lot of overlap with our problems: complex system, various horizons, rich data...
0
0
7
Something I wrote up w/ help from my team. There's a lot to say on this topic, and a lot of assumptions, hopefully this partially addresses some of them. Also a chance for me to shout-out
@beenwrekt
(this originally had "ImageNet generalize to ImageNet" too but had to pick one!)
This week on the HRT Beat, Algo lead Iain Dunning discusses the role of academic research in algorithmic trading. But how do we know what is useful and what is not? Read the full story on the
#HRTBeat
>>
#AcademicResearch
#MachineLearning
0
0
2
0
1
7
At this week's
@wehrtyou
AI Labs (HAIL) Paper Survey Club I presented "Bootstrap your own latent" by
@c_tallec
@KloudStrife
et many al. avoids negative example trap of contrastive & excellent ablation studies make for great paper. Presented at new office!
0
0
6
#JuliaOpt
JuMP helped Juan Pablo Vielma test ideas for extended reforms. for MIQCP with
@gurobi
&
@ibmoptimization
http://t.co/kP8JTGoB6P
0
3
5
Good morning
#NeurIPS2019
! Already in town? Come to the Finance+AI session at the Expo this afternoon, 2 pm, to hear talks by
@jpmorgan
,
@twosigma
, and
@wehrtyou
, followed by a panel.
0
0
6
New post: solving a combination lock puzzle with JuMP and
#JuliaLang
http://t.co/QqYqlOiMnm
#orblog
#orms
1
4
6
.
@JuliaLanguage
/
#JuliaOpt
tutorial for
@mitenergy
- demoing optimal power flow w/ JuMP+Interact.jl+GraphViz
#orms
http://t.co/iqyaPBgyHj
0
0
5
Simple + n-dimensional implementation of `
#numpy
.gradient` using
@JuliaLanguage
's nice indexing and parametric types
http://t.co/w3Qxi6ogyk
0
1
6
Found in
@ORCenter
: 1992 book about ANALYZE software to analyze LP structures (with floppy disk!)
#orms
http://t.co/Yxn9wr313s
1
3
6
Experimenting with a searchable
@JuliaLanguage
package listing - playing with
@angularjs
is addictive!
http://t.co/jmtcOTf10x
0
0
6
PhD students (and their supervisors!): the second HRT AI Labs Fellowship will be running summer 2020. Not a normal finance internship - we're aiming to do more basic AI research using all our data and all the compute you need! See for more info
@wehrtyou
1
1
6
New paper: "What Works Best When? A Framework for Systematic Heuristic Evaluation"
http://t.co/13A0lZo8Jo
#reproducibility
#heuristics
#orms
0
4
6
@ericjang11
Wrote a similar paper about heuristics for a classic NP-hard problem, MaxCut - no sign of progress over many years once you compare properly
0
1
6
#JuliaCon2015
#JuliaOpt
2.5 hour workshop is up! I'm @ 1hr:37, MIO for scheduling & a nonlinear rocket control model
0
1
5
Awesome graph layouts (usually!) using my
@JuliaLanguage
GraphLayout.jl pkg (w/
@danielcjones
's Compose.jl for viz):
0
4
5
My preview of
#JuliaCon2015
, June 24th-27th at
@MIT
's Stata Center:
http://t.co/KoN3swS4ot.
Register now at
http://t.co/XOaghnaF7M!
0
7
5
1 talk down, 1 to go
#informs2014
. great session with
@4er
@thyunes
John Siirola about subtleties and future of (MILP) modeling
#JuliaOpt
1
1
5
Spotted in SE Florida - NZ's craft beers
@MOABEER
,
@TuataraNZ
, and
@8wiredbrewing
- looking forward to tasting later!
http://t.co/DrLf3uVcA6
2
2
4
@lemergenz
In the 90s, who can say! Most techniques still have some use even today though (except maybe... SVMs?! Very rarely the solution to anything)
1
0
4
Our
@JuliaLanguage
for
#ORMS
/JuMP paper is available in
@INFORMS
Journal of Computing:
http://t.co/8Aq5GeUZ2K
please cite if you use JuMP!
0
3
5
New
#orms
paper using Julia+JuMP to solve stochastic programs
http://t.co/36RyvMNJXf
- learn more about
#JuliaOpt
at
http://t.co/z2bj6xzx7g
1
5
5
#JuliaOpt
JuMP 0.10 out now: vectorized ops, semidefinite programming, norm, loads quicker, see
#orms
@JuliaLanguage
0
5
5
My talk about
#JuliaOpt
from
#JuliaCon14
is up! video
http://t.co/8sZA647adq
+ slides
#orms
@JuliaLanguage
0
1
5
#JuliaOpt
JuMP (modeling optimization problems) will 'precompile' w/ next
@JuliaLanguage
release for speedy startup
1
3
5
An implication: recent multiagent RL research is mostly not applicable, as it can't handle N-party interactions between agents with a variety of unknown preferences in positive-sum interactions. Self-play is not even close to what you want or need.
0
0
5
@dellsystem
It's been pretty astoundingly to behold, apparently there is nothing SV cannot solve with Medium thinkpieces
0
0
5
Love
#JuliaOpt
? Love
#stickers
? DM me your name, address & quantity and we'll snail mail you some 2"x2" stickers!
http://t.co/52MoydW9my
1
0
5
@JuliaLanguage
0.4.0 is out now! Ann:
http://t.co/gSHnF8Xdm2,
upgrade today! (your code should work, maybe with deprecation warnings)
0
1
4
"linear programming came up in my machine learning course. it was described as a hill climbing algorithm.part of me died"
#justbigdatathings
1
1
4
0
1
4
New interview question? "if one chess GM
@MauriceAshley
plays 10
@wehrtyou
employees simultaneously, what is the expected number of turns until the first checkmate?"
0
0
4
Just presented about
#JuliaOpt
http://t.co/2MCaAMZuhZ
at
#JuliaCon
, modeling + solver interfaces to make
#ORMS
easy
0
1
4
Wrote a short blog post that explores the nature of the inter-package dependencies in the
@JuliaLanguage
ecosystem
http://t.co/7OyoD6dysk
1
4
4
New post:
http://t.co/W3Iz0nh9Ws
using the new cross-solver callback support in JuMP to solve TSPs using lazy constraints
#orms
#JuliaLang
0
5
4
Come to Miles Lubin's "JuMP: Nonlinear Modeling with Exact Hessians in
@JuliaLanguage
"
#informs2014
SD51, Sutter A, meet the
#JuliaOpt
team!
0
0
4
When picking prescriptive analytics/
#orms
tools, choose open source where possible, avoid vendor lock-in, especially if you are in academia!
1
1
3
New blog post: Analyzing
@JuliaLanguage
's
@github
issue counts over time
http://t.co/aNHYlaJGeZ
(using Julia, of course!)
0
0
4
New post: Sudoku-as-a-Service with
#julialang
-
http://t.co/VyZY5cl110
- MIP Sudoku solver exposed as webservice
#saas
#orblog
0
5
4
If you're talking about HRT and somewhere like Millennium as being fungible in any sense, either as types of firm or as out-of-college jobs... ngmi.
1
0
3
Acceptances going out today for talks at second
@JuliaLanguage
conference,
#JuliaCon2015
! Coming to
@MIT
this June
http://t.co/XOaghnaF7M
1
2
3
New post: solving an urban planning puzzle with JuMP and
#JuliaLang
http://t.co/RhGiJlDM4k
#orblog
#orms
0
3
4
@sasuke___420
This was one of our intern projects last year ("Signal Architecture Search"), was a lot of fun
1
1
4
Thanks
@polynoamial
for stopping by!
HRT’s speaker series is built to feed our culture of curiosity and is designed to bring a social aspect to learning and expanding our minds. Last week, Noam Brown, research scientist at Facebook AI Research, joined us for a fantastic virtual talk and Q&A session. Thank you Noam!
0
1
2
0
0
4
JuMP 0.1 is released - efficient MILP/MIQCQP modeling language
#orms
for
#JuliaLang
- docs: , see us at
#informs2013
!
0
3
4
At this week's
@wehrtyou
AI Labs (HAIL) Paper Survey Club,
@marckkhoury
(who just started last month!) presented "Hybrid Discriminative-Generative Training via Contrastive Learning" by
@lhaoml
@pabbeel
, with some great background (including a Vapnik quote, always a strong move)
0
1
4