All Weeks Python and Machine Learning for Asset Management Quiz Answers
Python and Machine Learning for Asset Management Coursera Quiz Answers
Week 1: Python and Machine Learning for Asset Management
Q1. What are some of the lessons from machine learning that can be applied to investment management? Multiple responses possible (2 correct answers)
- Wisdom of the crowd allows for a set of weak classifiers to be combined into a strong classifier
- Reduce overfitting by careful out-of-sample analysis (train and test)
- The selection of the best machine learning algorithm is always easy to determine since these algorithms provide performance guarantees.
- Machine learning algorithms are able to predict a stock market crash with 100% accuracy
Q2. What areas of FinTech that could be highly disruptive to the large banks and other financial organizations? Multiple responses possible (3 correct answers)
- Wealth management via robo-advisors
- Regulators will no longer need to be concerned about the survivability of banks to withstand a market crash. Thus, stress testing can be eliminated
- Insurance companies are able to drop re-insurance contracts since the worst-case risks can be avoided with perfect confidence
- Direct online lending customers
- Online insurance
Q3. How can machine learning algorithms improve investment performance?Multiple responses possible (2 correct answers)
- Predict all rare events with perfect accuracy, such as an upcoming pandemic
- Better estimates of the probability of a crash regime
- Implement automated re-balancing and tax-based strategies
- Predict changes in tax law in the future
Q4. What are the limitations of machine learning in investment management?
- Human behavior such as political risk disrupts the ability to predict future performance by machine algorithms
- Investment performance does not adhere to continuity conditions
- A majority of investors must employ automated investment systems in order to improve the performance
Q5. How can you tell if a machine learning method is successful?Multiple responses possible
- Evaluate ex post results
- Always select the model with the highest in-sample performance and lowest in-sample errors
- Ask the developers of the machine learning tools
Q6. Can machine learning algorithms be applied without human intervention to significant strategic decisions in which there are considerable uncertainties? True or false?
Q7. Do classification algorithms always provide probability estimates and false positive and false negative estimates as part of their normal output? True or false?
Q8. In the output of a graphical network analysis, what is the implication of two points that are close in terms of distance between them?Multiple responses possible (2 correct answers)
- No relationship via distance between nodes (objects)
- Nodes in the middle of the graphical network play an important connecting function in the network
- Nodes that are close to each other have stronger relationship than nodes that have large geographical distance
Q9. Reinforcement Learning has had a major area of research over the past few years. Which of the following applications are not likely to be improved by implementing reinforcement learning algorithms over the next several years?
- Long-term financial planning for individuals
- Identifying the best house to purchase for a middle-class family
- Optimal betting strategies for online betting on professional athletic events such as the Super Bowl in the US
- Perfect predictions of the location of earthquakes occurring in cities such as San Francisco
Q10. Supervised learning requires labelled data. Thus, for forecasting purposes, there must be enough representative data for training and validation steps
Q11. There has been a massive increase in micro-level data – at the level of individuals, e.g. purchases of products and services. This data can be coupled with macro-economic data, e.g., GDP and inflation, and with market data, e.g. stock performance to improve investment performance
Week 2: Python and Machine Learning for Asset Management
Q1. What is the motivation for employing risk-factors when constructing a financial risk management system? Multiple responses possible (3 correct answers)
- Factor investing can Improve diversification
- Factor investing eliminates the need to estimate the expected performance of asset categories and securities in a forward-looking portfolio model.
- Factor investing can assist with setting forward-looking parameters such as the expected performance of hybrid asset categories
- Factor investing is based on the premise that we can uncover and measure risks that are embedded in asset categories with multiple risks, such as high-yield bonds, private equity
Q2. What are the standard assumptions for applying the traditional OLS regression framework?Multiple responses possible
- Economic and causal relationship between the explanatory and the dependent variable
- Additive and linear impacts for each explanatory variable/factor
- Factors display regression towards the mean over time
- Relatively independent explanatory variables/factors
Q3. What might explain the difference in the factor loadings during crash economic periods as compared with normal economic conditions?Multiple responses possible
- Market contractions are relatively rare and can be largely ignored by most investors since they quickly return to normal conditions
- High volatility has a ripple impact on many market- based risks
- Crash periods display high contagion
- The flight-to-quality syndrome causes the correlation between equities and government bonds to become negative
Q4. How to address factors that have high correlation with other factors?Multiple responses possible
- Employ ridge regression to choose the most significant factors
- Combine correlated factor into an aggregate meta-factor
- Ignore this issue
- Eliminate one or more of the mostly highly correlated factors
Q5. How can you identify if you are missing a prominent factor?
- The out-of-sample errors from the regularized regression are very high
- The expected return estimates from the factor loadings fail to predict well the actual ex post returns over a short time period
- One of the popular factor models has not been working well over the past several months.
Q6. What are the disadvantages of factor approaches? (3 correct answers)
- Factor models are more complicated than traditional asset allocation methods
- Factor models require a two-step process when constructing an optimal portfolio. This process requires greater training and expertise than traditional portfolio models.
- The approach can be subjected to lower returns going forward to its popularity and overuse
- There is no universal acceptance for any particular set of factors and whether a portfolio of assets should be analyzed security by security or as a single time series
Q7. What are three of the major factors for high yield bonds?Multiple responses possible
- Global equities
- Political party in power
- Yield spreads
- Interest rates
Q8. Use the OLS Model under section 6, OLS of the notebook to answer this question. Suppose we increase the expected return of 10 Year US Treasuries by 1 percentage point. What would be the
change in expected commodity returns given the model?
- No change
- Increase expected commodity returns by .128%
- Increase expected commodity returns by 1.28%
- Increase expected inflation by 1.8%
Q9. Use the LASSO model under section 3, LASSO in the notebook to answer this question. This question covers the effect of changing lambda hat on a LASSO regression. Which of the following lambda hat values results in a LASSO model with all zero factor loadings, and which lambda hat value results in a LASSO model with no zero factor loadings.
- Lambda hat = .005 gives all zero factor loadings, and lambda hat = .000005 does not give all zero factor loadings.
- Lambda hat = .000005 gives all zero factor loadings, and lambda hat = .005 does not give all zero factor loadings
- Lambda hat = .005 gives all zero factor loadings, and lambda hat = .000005 gives all non-zero factor loadings
- Lambda hat = .000005 gives all zero factor loadings, and lambda hat = .005 gives all non-zero factor loadings
Q10. True or False: On average the factor loadings for a regularized regression will be less than the factor loading for OLS regression.
Q11. This question covers the procedure of cross validation. When we say “cross validation selects the “best” value of lambda hat.” What do we mean by “best?”
- This value of lambda results in the smallest mean squared error in the out of sample tests
- This value of lambda results in the smallest mean squared error in the training data
- This value of lambda results in the smallest difference between the mean squared error out of sample and mean squared error in sample
- It is smallest lambda that sets at least one factor loading to zero
Week 3: Python and Machine Learning for Asset Management
Q1. What are examples of successful applications of unsupervised learning in finance? Multiple responses possible (3 correct answers)
- Graphical network analysis to identify clients who might be amenable to new and novel services
- Indicate the conditions of a crash period
- Unusual behavior by credit card holders within a fraud detection system
- Identifying firms with unusual levels of income for their respective industry and investments
Q2. Can PCA be applied to the time series of a portfolio of stock prices?
Q3. What is the optimal number of clusters for a particular set of stocks such as the S&P 500? Multiple responses possible
- There is a tradeoff between the number of clusters and the degree of parsimony in a cluster.
- The cluster analysis should provide summary statistics on the similarity measure for stocks within a cluster and chose the number of clusters that minimizes similarity
- Clustering output requires at least 10 clusters.
- The analysts should plot the total similarity measure as a function of the number of clusters and select a point on the elbow of this curve.
Q4. How might you identify a firm that has highly unusual behavior from the output of a cluster analysis? Multiple responses possible
- The stock would appear in its own cluster
- Stocks in a single industry will always show up in a common cluster
- Stocks with similar behavior should be identified in a single cluster
Q5. Much of the success in machine learning is due to access to massive data on customer behavior coming out of user tracking. True or false?
Q6. Can unsupervised learning make decisions about the best stock to invest in? True or false?
Q7. What did the empirical tests show regarding the application of PCA/clustering and via graphical analysis to a stock selection application?Multiple responses possible
- The graphical analysis was slightly improved over the PCA/clustering method
- Both methods outperformed the Markowitz portfolio
- The empirical tests did not improve performance
Q8. Does the graphical analysis fit in the area of supervised learning? True or false?
Q9. In the discussed empirical tests between the PCA/clustering algorithm and the graphical network analysis, did one of the approaches dramatically outperform the other? True or false?
Q10. Most clustering methods employ heuristic algorithms, as compared with a formal optimization model. What leads to this situation?Multiple responses possible
- The optimization model is much harder to solve, especially for an application with a large number of points/customers/variables to cluster
- The heuristic methods are easy to interpret
- The distance function and objective function required by the optimization presents an informational barrier
Q11. A cluster always has at least two stocks
Q12. Which of the following sentences correctly summarizes the relationship between the graphical network and the number of sectors?
- The number of clusters must not be grater than the number of sectors
- The number of clusters must be grater than the number of sectors
- The number of clusters and the number of sectors must be equal
- None of above
Q13. When graphical analysis is performed, the results are different for different lengths of time (6M, 1Y, 5Y). What are the differences and what could cause these differences?
- The longer the length of time, the more separated the clusters: using more information the trend of how stocks behave is clearer and then companies that are not in the same cluster show more discrepancies in their trends.
- The longer the length of time, the more separated the stocks: using more information the trend of how stocks behave is less clear and then companies show more independent trends.
- The shorter the length of time, the more separated the clusters, using less information the trend of how stocks behave is clearer and then companies that are not in the same sector show more discrepancies in their trends.
- The shorter the length of time, the more separated the clusters, using less information the trend of how stocks behave is clearer and then companies show more independent trends.
Q14. In this notebook, when we calculate for the summary statistics, we notice that the stock “BABA” was dropped because it has missing data for our specified period of time. Please revise the notebook to display the summary statistics
information for “BABA” over the full period of time for which its data are available. What are the “Annu. Ave Return” and the “Annualized Sharpe” for
“BABA”? (Hint: the command “first_valid_index()” can be useful here.)
- 21.77%; 0.63
- 19.47%; 0.72
- 11.26%; 0.66
- 8.84%; 0.45
Q15. Using the notebook, perform graphical analysis with S&P500 (i.e. 23 stock returns in total) for the time period 2015-07-01 to 2020-06-28. Which of the following sentences correctly summarizes the resulting network plot? Multiple responses possible.
- *A: 4 clusters are identified
- All banks are identified to be in the same cluster
- APPL, GOOG, MSFT, and SP500 are identified to be in the same cluster
- HSBC and JNJ are not connected directly by an edge
Q16. In this notebook, provided that two stocks in the network graph are disconnected (i.e. are not connected directly by an edge), what can we conclude about the relationship between these two stocks?
- The two stocks are independent of each other
- The two stocks are independent conditionally on the others
- The two stocks are in the same cluster
- The two stocks are from the same sector
Q17. Using the notebook, perform graphical analysis without S&P500 (i.e. 22 stock returns in total) for the time period 2015-07-01 to 2020-06-28. Which of the following sentences correctly summarizes the resulting network plot? Multiple responses possible.
- 4 clusters are identified
- All banks are identified to be in the same cluster
- APPL, GOOG, MSFT, and BABA are identified to be in the same cluster
- HSBC and JNJ are not connected directly by an edge
Q18. Comparing the two network plots for the graphical analysis with and without S&P500 (i.e. 23 and 22 stock returns in total, respectively) for the time period 2015-07-01 to 2020-06-28, which of the following observation is correct? Multiple responses possible.
- Both analyses group the stock returns into 4 clusters
- BABA is identified to be in a cluster of size one, regardless of whether or not S&P 500 is included
- The international banks (HSBC, RY) tend to be positioned at the periphery of the bank sector, regardless of whether or not S&P 500 is included
- None of the above
Week 4: Python and Machine Learning for Asset Management
Q1. What are the characteristics of a crash regime? Multiple responses possible
- Sharp drop in prices across many markets
- Layoffs in the financial sector
- Higher transaction and market impact costs (e.g. bid/ask spreads)
- High volatility
- Higher correlation than normal regimes (contagion)
Q2. What are the critical inputs of a regime-switching financial risk management system? Mutiple responses possible (3 correct answers)
- A transition matrix for the probabilities of movement from time t to time t+1
- The state of current regime
- Knowing the exact sequence of regimes for the future
- A mechanism for addressing multiple and conflicting objectives
Q3. What are the assumptions for converting a multi-period investment model into a sequence of single period models? Multiple responses possible (3 correct answers)
- No transaction and market impact costs
- Independent and identically distributed returns at each period, including temporal independence
- No inflows or outflow of cash or liabilities (asset-only allocation)
- Lack of expertise to construct a multi-period model
Q4. What are a few of the reward and risk measures that apply in a multi-period long-term investment system? Multiple responses possible
- Probability of meeting a target goal at the horizon date T.
- The expected returns (or the median return) at several key junctures: 1 year, 3 years, 5 years, and beyond
- Downside risk relative to a goal
- Standard deviation of the key parameters –including wealth at the horizon T.
- Goals-at-Risk and Conditional Goals-at-Risk
Q5. An advantage of the multi-regime model is the greater realism of the forecasting system as compared with a single regime model: Yes or no ?
Q6. The scenario based portfolio model is identical to the traditional mean/variance (Markowitz) model under which of the following assumptions?Multiple responses possible
- The scenarios are derived from a multi-normal distribution (single regime)
- The number of scenarios is large enough to reduce the sampling errors to a management number.
- The scenarios are based on equal periods of historical performance.
Q7. What are the advantages of a multi period regime model over the traditional Markowitz model?Multiple responses possible
- More accurate future projections compared with historical performance patterns
- Ability to include inflows/savings and outflows/expenses within the model directly
- Ability to address transaction costs
- Ability to model temporal risk measures, such as probability of meeting a goal at a designated future date
Q8. In the Module 4 Jupyter notebook, please kindly run the Q-Q plot function for each asset. Which asset has a similar tail to the normal
- High Yield
- US Treasuries
- US Equities
Q9. In the Module 4 Jupyter notebbok, please kindly set the random seed as 777. Simulate 8 assets with 10000 scenarios, each containing 600
monthly returns (50 years of asset returns). What is the (arithmetic) mean of all simulated
returns of US Treasuries?
Hint: The average is taken from 6000000 values in total. Since we ask for return, you should
subtract 1 from the simulated values.
Q10. What are considerations in calibrating the parameters for a scenario generator ? Multiple responses possible (3 correct answers)
- Expert judgment
- Current economic conditions such as interest rates and spreads
- Historical performance
- An AI expert system that provides the needed information
Week 5: Python and Machine Learning for Asset Management
Q1. What are the primary assumptions when estimating a crash regime? Multiple responses possible (3 correct answers)
- The underlying causes are the same for each crash
- Discovering factors that have legitimate reasons for repeatability
- Availability of adequate historical data on factors and market based asset performance
- Ability to apply a non-parametric method, such as support vector machines
Q2. What can be done with better probability projections – Multiple responses possible (3 correct answers)
- Provide details to the central bankers (head of Federal Reserve in the U.S.) so that they change policy to avoid the crash
- Nothing since most investors are too cautious to take action when warned
- Employ within a strategy at the casino for betting on slot machines
- Incorporate into the scenario generation of a multi-regime investment model
Q3. What are the primary assumptions of the “wisdom-of-the-crowd” approach?
- Combining weak learns in a aggregation will lead to improved results as compared with a single learner
- The voting process provides a democratic approach to decision making
- The crowd must be experts with doctoral degrees in the target domain
Q4. How can regularized regression LASSO assist with the selection of explanatory variables? Multiple responses possible
- The out of sample testing, via cross validation improves the robustness of the betas
- Adding penalty terms shrinks the beta coefficients and thereby reduces the number of non-zero betas
- The regularization algorithm always selects a single beta factor, the most important feature
Q5. Many economic time series have relatively high correlations. How can a forecasting system based on machine learning address this issue?Multiple responses possible
- Ridge regression can improve the condition number for the resulting system of normal equations
- Regularized regression shrinks the beta coefficients and reduces the number of non-zero factors
- Highly correlated explanatory variables is not an issue for traditional regression and similar econometric models
Q6. What is a reasonable limit on the number of explanatory variables in a feature selection study?
- Five hundred to a few thousand
- Five billion
Q7. Identifying a crash period requires judgement about the size and extent of the negative performance. Multiple responses possible
- Crashes are relatively rare events
- It is relative easy to see a crash on the horizon
- Modern machine learning algorithms give us a full proof forecasting system
- Crash periods entail sharp losses and drawdowns
Q8. In this notebook, we are building forecasts for recessions with macroeconomic indicators.
What type of problem is this in Machine Learning?
- Unsupervised Learning – Clustering Problem
- Supervised Learning – Regression Problem
- Supervised Learning-Classification Problem
Q9. Is this data balanced or imbalanced?
Q10. Why a special type of Cross Validation (TimeSeriesSplit) is used in this notebook?
- Time series structure in the dataset
- Cross-section data
- Panel data
Q11. ROC metric is used to evaluate performance of the prediction algorithms for our problem.
We plot the ROC curve for the best performing ones. What are the axis’s of the plot
- True Positivity Rate against False Negativity Rate
- True Positivity Rate against False Positivity Rate
- True Negativity Rate against False Negativity Rate
- True Negativity Rate against False Positivity Rate
Q12. Which algorithm performed best in predicting recessions according to results in the note-
- Logistic Regression with penality
- Decision Trees
- Random Forest