Skip to content

Literature Review

Meta

  • Year - Title
  • Paper authors:
  • Link to the paper (ideally on gdrive)
  • Review author / date
  • Score in [0, 5], where:
  • 5/5: Must-read
  • 4/5: Some interesting ideas we can reuse
  • 3/5: Pretty much what one would have done as first experiment
  • 2/5: ...
  • 1/5: Horrible: same bullet-proof logic as in a politician speech
  • Summary:
  • At most 5-10 bullet points explaining what the paper tries to accomplish
  • Describe the data used, setup, model formulation, ...
  • Good references
  • Praises:
  • At most 5 bullet points
  • Focus on what is different, interesting, and not on the obvious
  • Critiques:
  • At most 5 bullet points
  • Explain what is not solid in the analysis, suggestions on how to improve
  • Next steps:
  • What next steps should we take, if any, e.g.,
    • Read the bibliography
    • Try experiments

To cut and paste


### Year - Title
- Paper authors:
- [Link]()
- Review author / date:
- Score:
- Summary:
- Praises:
- Critiques:
- Next steps:

News for commodity prediction

2015 - The role of news in commodity markets

  • Paper authors: Borovkova
  • Link
  • Review author / date: GP, 2019-11-22
  • Score: 4/5
  • Summary:
  • Dataset: prepackaged Thomson-Reuters sentiment (TRNA)
  • Studies the effect of sentiment on commodities through event studies
  • Forecast prices and volatility
  • Praises:
  • Decent statistics about the data set
  • States that one needs to understand if the sentiment is attached to demand and supply
    • Not sure if TR actually does that
  • Confirms our point about "momentum-related news" (i.e., news about the fact that the price is going up)
  • Confirms periodicity we are aware of
  • Interesting local level model to extract the hidden sentiment
    • Very similar to what we thought to do (including the idea of using Kalman smoother)
  • Critiques:
  • Nothing really
  • Next steps:
  • Understand if TR considers sentiment distinguishing supply or demand
    • We should do this (not sure how PR does that)
  • Remove carefully momentum-related news
  • Remove or count carefully repeated news (maybe use a measure of similarity between articles)
  • How to deliver "event study" models to customers? Should we "unroll the model" for them providing a stream of predictions?

Social sentiment

2015, Predicting global economic activity with media analytics

  • Paper authors: Peterson et al.
  • Link: In Tech/papers
  • Review author / date: GP, 2019/12/08
  • Score: 2/5
  • Summary:
  • Predict PMI indices (which are related to the
  • Praises:
  • Interesting approach for going beyond polarity in sentiment considering
  • Critiques:
  • No seasonal component
  • Usual problems with methodology OOS
  • Next steps:
  • Consider the TRMI "indices" (optimism, fear, joy, trust, violence)
  • Consider the difference in professional news vs social news sentiment
    • What does it mean if there are large statistically significant difference?

2018 - Twitter, Investor Sentiment and Capital Markets, what do we know?

  • Paper authors:
  • Review author: GP, 2019-08-21
  • Link:
  • Score: 3 / 5
  • Summary:
  • Good survey of the literature about social sentiment used for finance
  • Most authors report predictivity of social sentiment for:
  • Different metrics (returns, risk, trading volume)
  • Assets (US stocks, exchange rates, commodities)
  • Events (IPO, earnings)
  • Next steps:
  • Read all the bibliography and reproduce some of the results
  • TODO: Update this to new template

Time series

On-Line Learning of Linear Dynamical Systems: Exponential Forgetting in Kalman Filters

  • Paper authors: Mark Kozdoba, Jakub Marecek, Tigran Tchrakian, and Shie Mannor
  • Review author: Paul, 2019-12-02
  • arXiv, AAAI
  • Score: 4/5
  • Summary:
  • Interesting insight into how to approximate a non-convex optimization problem with an approximate convex one
  • Shows that for observable Linear Dynamical Systems with non-degenerate noise, the dependence of the Kalman filter on the past decays exponentially
  • For this class of systems, predictions may be modeled as autoregressions. In practice, not many terms are needed for a "good" approximation.
  • The algorithm is on-line
  • Comparison to the Kalman filter is formalized with regret bounds
  • IBM / Technion research
  • The setting is one where we are learning the best fixed but unknown autoregression coefficients (rather than one where we are interested in truly dynamic updates)
    • The learning rate decays like $1 / \sqrt{t}$, and so under some mild constraints on the time series being modeled, the autoregression coefficients converge
    • The linear dynamical system setup considered is one where the state transition matrix and the observation direction are time-independent
  • Praises:
  • References standard big works in the time series literature, like West and Harrison (1997) and Hamilton (1994)
  • Introduces a relatively simple online technique that competes well with the more complex Kalman filter
  • Critiques:
  • Bounds / constants aren't quantitative
  • Next steps:
  • Look at the code accompanying the paper: https://github.com/jmarecek/OnlineLDS
  • Implement and compare to, e.g., z-scoring (a particularly simple case of Kalman filtering)
  • If we have a long history, it may be better to perform a single autoregression over the whole history
    • This suggests
  • What if we keep the learning rate fixed over time?
    • This would effectively allow for "drifting" dynamics
    • The proofs of the results of the paper would no longer apply
    • It isn't obvious how the learning rate ought to be chosen

Predictive State Smoothing (PRESS): Scalable non-parametric regression for high-dimensional data with variable selection

  • Paper author: Georg M. Goerg
  • Review author: Paul, 2019-12-03
  • Link
  • Score: 4/5
  • Summary:
  • A kernel smoother, but unlike traditional ones, it
    • Allows non-local (with respect to the x-var space) pooling
    • Is scalable (e.g., computationally efficient)
  • PRESS is a generative, probabilistic model
  • States are interpretable
  • Compatible with deep neural networks (though experiments referenced in the paper suggest depth doesn't help, e.g., a wide net with one softmax is enough)
  • Competitive with SVMs, Random Forests, and DNN

  • Predictive state representations are statistically and computationally > efficient for obtaining optimal forecasts of non-linear dynamical systems > (Shalizi and Crutchfield, 2001). Examples include time series forecasting > via epsilon-machines (Shalizi and Shalizi, 2004)...

  • Praises:
  • Combines some clever insights
  • References a TensorFlow implementation and suggests that implementing in various frameworks is straightforward
  • Critiques:
  • No pointers to actual implementations
  • Time series applications are referenced in Section 2, but many relevant (to our work) practical ts-specific points are not developed in the paper
  • Next steps:
  • See if someone has already implemented PRESS publicly
  • If no implementation is available, scope out how much work a minimal pandas-compatible implementation would require

2019, High-Dimensional Multivariate Forecasting with Low-Rank Gaussian Copula Processes

  • Paper authors: David Salinas, Michael Bohlke-Schneider, Laurent Callot, Roberto Medico, Jan Gasthaus
  • Review author: Paul, 2019-12-28
  • arXiv
  • Score: 4/5
  • Summary:
  • Learns covariance structure and model together
  • Handles series with time-varying, high-dimensional covariance structure
  • Simultaneously handles series at different scales (in terms of the range)
  • Uses a non-linear, deterministic state space model with transition dynamics parametrized using an LSTM-RNN
  • Praises:
  • Implemented in GluonTS (https://github.com/awslabs/gluon-ts/pull/497) by one of the coauthors who works on time series forecasting at AWS
  • Code for the paper at https://github.com/mbohlkeschneider/gluon-ts/tree/mv_release
  • Good choice of baselines comparisons
  • Demonstrates the importance of data transformations
  • Next steps:
  • Use in cases where we have a large number of time series known to have meaningful correlations

2014, The topology of macro financial flow using stochastic flow diagrams

  • Paper authors: Calkin, De Prado
  • Link
  • Review author / date: GP, 2020-01-17
  • Score: 1 / 5
  • Summary:
  • Praises:
  • PCA on futures sectors
  • Interesting graphical representation
    • Width of the arc represents strength of relationship (in terms of $R^2$)
    • Color (green / red) and intensity represent sign and magnitude
    • Lags are delays
    • Geometric topology represents relationships better than tables
    • Connectivity of a vertex represents importance
  • Agreed that econometrics as it is, is close to a pseudo-science that more complex techniques are needed than inverting a matrix
  • Critiques:
  • Various inflammatory remarks and very little content
  • Next steps:
  • None

Computer engineering

2015, Hidden technical debt in machine learning systems

  • Paper authors: Sculler et al.
  • Link
  • Review author / date: GP, 2020-01-07
  • Score: 4/5
  • Summary:
  • Many interesting little observations about ML practices and engineering
  • Praises:
  • Validates how approach of minimizing technical debt and paying it off the interest, e.g.,
    • Treat configuration as code, as we do
    • Design abstraction carefully
    • Routinely clean up the code
    • No distinction in quality between research and production
    • Use a single language for everything
    • Need for committing to the healthy engineering practices
  • Critiques:
  • None
  • Next steps:
  • None