Experimentation & Testing (Notes from March 25 - March 31, 2019)

I know, I know. It's a bit of a cop out to use a Game of Thrones image on the back of the Season 8 premiere from Sunday. Sue me [please don't]. And I'll give credit to the image creator: Instagram @chartrdaily for the fun visualization. However, after listening to Pinnacle Sports' Marco Blume, I couldn't help after hearing deployment strategies for their prop bets on popular TV shows, such as who will be left on the Iron Throne or the ever popular "Who dies first?" props. They experiment, hypothesize, post a line with a limit (hedge risk) and let the market decide from there. And boom - we have the theme of the week!Antoine Nussenbaum, of Felix Capital at the time, mentioned going from private equity to start-ups and venture funding where they had to decide between backing people or belief in the company. He got first-hand experience by starting a company with his wife, successfully gaining funding, and then exiting - only to fail with a different company that wasn't scaling. How did he go through frameworks to decide on startups to fund or help?Mark Suster gave his take on how he comes to investment funding - sales, technical skills and being aware of each. How did his entrepreneurship experience influence his framework for funding new start ups? Why is it that there is a sweet spot for amounts based on run rate? Experimenting, failing and adjusting.Then I had listened to 2 data scientist / researchers in their discussions of NLP parts - what to test, what they assumed to be true, how to approach new methodology and testing this methodology. Is there a limit to the progression that can be made with NLP? Why might it be relevant to decide on testing state-of-the-art further? Then, ultimately, what's the applications for how we can use that optimization to improve the current status quo?I hope everyone checks out what may interest them - this was a fascinating and fun week. So much so, that I suggested to a few different students for them to check out different parts (granted, I do this often, but I was quite excited to share these ones).Cheers!

  • Antoine Nussenbaum (@Nussenbaum), Principal and cofounder of Felix Capital (20min VC 084)
pvcmh-1_
  • Partner at Atlas Global prior, p/e fund that was part of GLG Partners
  • Working on digital early-stage, venture fund and helped startups bootstrap after missing the tech side
  • Miraki, Jellynote, Pave, Reedsy, and 31Dover as some of his best investments
  • Helped start Huckletree with his wife Looked for investment of $80mln but got $120mln
  • Backing someone vs backing the company initially in early stage funds
  • Raised in Paris in international environment, lived in UK as well
  • Launched 2004 software-on-demand business with 2 friends "that was not scalable at all"
  • Did M&A in the UK after leaving software
  • Felix Capital at intersection of creativity + technology, lifestyle brands: ecommerce and media, enabling tech
  • Stages - flexible capital, but have made investments from $200k - $6mln, focus on Series A + B
  • Geographic - agnostic, as long as backing entrepreneurs
  • Advisory services and focused on helping their investment companies
  • More entrepreneurs that know the playbook and how they can build, grow and scale Looking for more companies that can scale globally or expanding outside with proper funding
  • Using Triangle as an example - bathing suits on Instagram strategy and launching millions of product via digital
  • ProductHunt as a blog he gets lost in - 15 min of destruction
  • Lifestyle-related excitement: food side, better life, marketplaces
  • Hard Thing about Hard Things and Capital in the 21st Century - relationship of wealth and economic wealth
  • Mark Suster (@msuster), MP @ Upfront Ventures (20min VC 085)
8647fd890a54e10bd320ada2651040c5
  • Was VP of PM at Salesforce.com before Upfront
  • Late 80s - had an interest in development as a student in college in the UK
  • Worked initially as a programmer at Anderson (Accenture) for 8 years
  • Entrepreneurship isn't for everyone - better to start earlier, need to have a fundamental understanding of systems (coding)
  • Python, PHP, Ruby, JavaScript - not trying to become best developer - just knowing the systems
  • Sales experience would be second - telesales or customer support - ask CEO to do an hour a week of calls
  • Started 2 software companies - one in England and then Silicon Valley, selling both - backer brought him in to VC Fred Wilson wasn't an entrepreneur, but does give you the insight
  • Don't get the sense of urgency with too long a time - 3 months vs 12 months
  • Too much capital creates laziness and shortcuts that lead to mistakes
  • 18 month run rate for capital - takes 3-4 months to raise (start with 6 months plus)
  • Wants to see early stage companies once a month, roughly.
  • $240mln fund - invest half into companies and reserve the other half for follow-ons 3 year timeframe, $40mln with 5 partners - $8mln per partner Series A, B rounds where each partner is doing 2-3 deals per year when avg is $3-5mln investment
  • On his blog, has the "11 Attributes of Entrepreneurs"
  • Best known post would be "Invest in Lines, not Dots" - x-axis as time, y-axis is performance (any given day, your dot) Interactions create a line that matches a pattern and he can decide if he wants to do business
  • Not a big fan of deal days or investor days where you hype up a company because of this
  • 50 coffee meetings a year - once a week, if you meet 50 entrepreneurs a year, maybe you'll become close with 5-10 of them Single best introduction is from a portfolio company CEO for an investor
  • He knows and built software company - SaaS-space since he knows how to be helpful
  • Data and video tech industry (has 11 personal investments and 5 are video)
  • AgTech as an underappreciated industry so far - stays quiet until a few investments before hyping
  • Too much company, too much money and entrepreneurs clouding the market for everyone else
  • Book "Accidental Superpower", how demographics and topology will drive the future and how areas grow
  • Marco Blume, Trading Director at Pinnacle Sports (DataFramed #54 2/18/19)
pinnacle_logo
  • Got into data science by "sheer force", building quant team out from Excel going to R
  • Efficiency was by orders of magnitude since R was better than Excel
  • Could do anything with risk management, trading, sports
  • Pricing GoT, hot dog eating contest, pope election and making the lines Use pricing and market analytics to let the people set prices
  • Risk management in general - maximize probability and hedging risk Does the bottom line change? Does it affect anything? Regulations.
  • NBA where all teams have played each other - have a good idea of strength of teams
  • Soccer or world cup - not as much certainty with teams not always playing each other
  • Start of season has a lot more volatility and responsiveness to bets because of uncertainty By end of season, bookmarkers have the price and knowledge, so they're likely to increase risk
  • Bayesian updating
  • Goals to improve models, open new betting options to clients
  • Low margin, high volume bookmaker - little bit with a lot of options
  • Book of Superforecasting - group of people who are better at forecasting Pays them already at Pinnacle - consultants, betting and paying the price
  • Much bigger R shop than Python at Pinnacle, active in the R community
  • R becoming more of an interfacing language and production language (vs C# or other), can use R-keras or plumbr
  • Teaching dplyr, rmarkdown and ggplot cover 95% of their work outside of specialists
  • GoT as one of his favorite bets
  • Matthew Peters (@mattthemathman), Research Scientist at AI2 - ElMo (Data Skeptic 3/29/2019)
ai2-logo-1200x630
  • Research for the common good, Seattle, WA research
  • Language understanding tasks - ELMo (embeddings from Language Models)
  • PhD in Applied Math at UW, climate modeling and large scale data analysis Went to mortgage modeling, tech industry with ML and Prod dev in Seattle
  • Trying to solve with very little human-annotated data, technical articles or peer-reviewed Very difficult, very expensive to annotate - can you do NLP to help?
  • Word2vec as method for text to run ML on text, context meanings of say, bank
  • ELMo as training on lots of unlabeled data
  • Given a partial language fragment, language modeling predicts what can come next
  • Forward direction or backward direction (end of context), neural network architecture
  • Research community may want to use ELMo, commercial use to improve models already in prod Pre-trained models available and open source
  • In the paper, evaluated NLP models on 6 tasks - sentiment, Q&A, info extraction, co-reference resolution, NL inference
  • Got significant improvements on results from the prior state-of-the-art models
  • Character-based vs word approach Single system should process as much text as possible (morphology of the word, for instance)
  • Paper over a year old now but Bert was put up on ArXiv to improve upon ELMo (transformer architecture for efficiency)
  • Scaled the model that could be trained by many X's, quality is tied to the size / capacity
  • Language modeling loss changed, as well (word removed from middle of sentence and predict before/after)
  • Large Bert models have computational restrictions - how far can you get by scaling the model
  • Kyle and early Data Science Hiring Processes (Data Skeptic 12/28/18)
github-logo-und-marke-1024x768
  • Success isn't correlated with ability to give good advice
  • Conversion funnel for businesses: website that sells t-shirts, for instance
  • Tons of ways to bring people into the door / website (ads, social media campaign, ad clicks)
  • Register an account or put into cart (what %, track it, a/b test and improve)
  • Cart to checkout process (how many ppl? Credit card entered, goes through, etc…)
  • Do any sites convert faster than others? Keep track, find out why / focus on continuing it
  • Steps for job hire: video chat / task / phone screens / on-site next / offer
  • Resume should be pdf (doc may not open nicely on Mac or otherwise) - include GitHub
  • SVM - should have margins or kernel trick on resume (otherwise, don't include it) Â Ex: ARIMA (auto-regressive integrated moving average) - time series data