How to Become a Data Scientist (Part 2/3)

Having read Chapters One and Two (i.e. Part One), you should now have a good comprehension of what commercial data science entails, the different forms it takes, and what is required to be a success in the profession. And having thought deeply about your motivations, you should have a clear picture of your goals, and ultimately – the type of data scientist you want to become. So give yourself a pat on the back, because you are now ready to begin the real fun: learning.

In this chapter, we will explore the options at your disposal – but first – we will begin proceedings by discussing an important notion that concerns data science and learning.

Continual Learning

Just like a doctor has to stay abreast of medical developments, learning never stops for a data scientist. The field (and the technology) is evolving so quickly; what you learn now might not be relevant in the years to come. Look at the rise of deep learning, to take just one example. This is what Sean McClure was alluding to in his post emphasising the importance of problem solving (highlighted in Chapter One).

Quite simply, if you are not passionate about the field and do not enjoy learning, then data science is not for you. Conferences and networking with the data science community are effective ways of keeping on top of the latest developments. And regularly reading books and papers is very important (on this: if you do not have a research background, it is worth learning how to read academic papers properly).

Play. Build. Experiment.

Going back to the message we touched on in Chapter One, there is only one-way to develop your capability as a data scientist: experience, experience, experience. I could launch into a lengthy discussion on this, but I happened to come across two excellent posts that cover the points I wanted to make, so have a read of Brandon Rohrer: A One-Step Program for Becoming a Data Scientist and Rossella Blatt Vital: The Scary Rise of the 'Fake Data Scientists'.

This is what should you take from these: data science is an expert field, it takes a long time to master, and you will only do so through practical experience. As James Petterson summarised:

“Nothing beats experience. You can read as much as you want, you can do all the Coursera courses, but unless you get your hands dirty, you won’t learn”

The good news is there are some great avenues to gain practical experience, and we will turn our attention to these now.  

Kaggle / Open-Source / Freelancing

If you haven’t heard of Kaggle, Google it... NOW! Kaggle is an incredible platform where you can play around, develop your expertise and learn, of course. James put it this way:

“If I hadn’t competed in Kaggle competitions, I would have finished my PhD without knowing the tools that people use in industry. For example, a lot of the methods used in industry are based on ensembles or decision trees, like random forests. They are really powerful and are my first choice in both competitions and industry, but I wasn't exposed to them during my PhD”

There you have it: you can improve your skills while learning the techniques that are commonly applied in industry. And if you start doing well in the competitions, it provides evidence of your capability, as we will see in Chapter Four.

Outside of Kaggle, another option is to contribute to open-source projects. A simple search on GitHub should reveal some projects you can start to sink your teeth into, and gain practical experience while doing so.

Finally, if you can get freelancing work, this is a great way to build a track record and demonstrates that you can operate in a commercial environment. And rather conveniently, you could even utilize the Experfy platform for that purpose.

To PhD or not to PhD

Do you need a PhD to be a data scientist? Not necessarily, but there are many advantages, as Sean Farrell noted: 

“The process of obtaining a PhD is a filter for creative problem solving skills [and it] shows you can master a particular field in a short space of time and become a world expert, which proves you’ll be able to do it again and again”

And apart from anything, it provides you with the time to study and to develop your skills. Furthermore, if you are interested in specialising within a specific area like image processing or natural language processing, then PhD research is certainly worth considering.

But going down this path is not the only way to data science. James did a PhD in Machine Learning (focused on researching a very specific type of method) and he feels that a lot of PhD research is not always applicable to industry, i.e. if your job is to apply machine learning rather than research it, you don’t necessarily need a PhD. As such, I asked him whether he thinks people should choose a PhD based on its relevance to industry and he said:

“If possible, but that’s really hard because most of what we do in industry is not state of the art, we use methods that have been around for years and apply them to different problems. There are exceptions of course: you might work at Google in research, for example. But most of the knowledge I use day-to-day, I learnt working [at Commonwealth Bank] and by competing in Kaggle. Of course, doing a PhD, you learn about the whole process, spend a lot of time doing experiments and learning how to do them properly, and that is valuable. But I wonder if you could learn that from other means?”

Given the right motivations and armed with an informative guide on how to become a data scientist (where could you find one of those I wonder?), I have no doubt it is possible to learn by yourself. But it is worth making the point again: there are no shortcuts; it requires a lot of self-study and getting your hands dirty – whatever path you take.

There is also the employability aspect to consider: are you more employable as a PhD graduate vs. spending the same time on self-study? I do not have sufficient evidence to comment, but either way, it is more important whether you have truly spent the time building up expert capability (and how you can evidence this). PhD’s are certainly valuable but there are great data scientists with PhD’s and great ones without.

Other University Degrees 

So a PhD is not for you – perhaps it is the cost, or perhaps you have not yet developed the expertise necessary for research of this nature. Whatever the reason – there is no need to panic – because many universities are now offering Bachelors, Masters and Diplomas specifically designed for data science, where both computer science and mathematics/statistics are on the curriculum (the attentive reader will remember this from Chapter Two). 

Courses like these will certainly take you in the right direction, but take note: they won't be enough to convert you into a ready-made data scientist, because as we know – that takes experience. 

Online Courses

In a similar sense – even if you come from another quantitative field – a few online courses will not make you an expert, and remember: this is an expert field. But even if an online course was enough to master a chosen subject, you will still face competition, who – in all likeliness – will have far more practical and commercial experience in these areas. This is really important to be conscious of, and so we will return to this in Chapter Four. 

All this being said, online courses are incredibly useful tools to help kick-start your journey, or begin learning a new area (like deep learning, for example). The most popular courses are found via Coursera, Udacity and edX, with Dylan Hogg describing Andrew Ng’s Machine Learning on Coursera as “an absolute pre-requisite for anyone who does not have a research background”.

The following is by no means a complete list, or mandatory for that matter, but these also stood out to Dylan and some of the other data scientists we have met so far:

• Machine Learning: Intro to Machine Learning (Udacity)
• Deep Learning: Deep Learning (Udacity), Neural Networks for Machine Learning (Coursera)
• Spark: Big Data Analysis with Spark (edX), Distributed Machine Learning with Spark (edX)

During my interactions with the Experfy team, I've found out that they were also launching a training platform. You can see a preview here. 

Books

Needless to say: good books are an invaluable resource and our favourite data scientists advocated the following:

• Pattern Recognition and Machine Learning by Christopher Bishop
• Machine Learning: a Probabilistic Perspective by Kevin P. Murphy
• Why: A Guide to Finding and Using Causes by Samantha Kleinberg (if you want to know why this is important, take a look at Yanir Seroussi’s blog post on: Why You Should Stop Worrying About Deep Learning and Deepen Your Understanding of Causality Instead)
• An Introduction to Statistical Learning by James, Witton, Hastie and Tibshirani, which, according to Dylan: “is a great introduction to statistical learning and is an accessible version of the more advanced classic”: Elements of Statistical Learning
• And for a different suggestion, Will Hanninger recommended The Pyramid Principle by Barbara Minto. It does not cover data science specifically, but is valuable for problem solving and presenting

Presenting / Communicating

If you do not have a natural disposition for communicating – especially with non-technical people – this is something you will need to work on (see Chapter One for why). Gaining practice and obtaining feedback is the best way to improve your soft skills, although Yanir also recommended the classic book by Dale Carnegie: How to Win Friends and Influence People.

HOW THE MODEL WORKS

At a high level, our approach is as follows. First, we estimate a regression model to predict the number of goals scored by a particular team (“team i”) against a particular opponent (“team j”) using the entire history of mandatory international matches since 1958, when the first European championship was played (a total of 4,719 matches).[1] Following the literature on predicting football matches, we assume that the number of goals scored by team i is described by a so-called Poisson distribution and explained by the following statistical factors:[2] 1. The difference in team performance as reflected in Elo ratings prior to the match. The Elo system was originally devised to rank chess players. It is a composite measure of national football team success that evolves depending on a team's results and the strength of its opponents. 2. The number of goals scored by team i in the last 10 competitive matches. 3. The number of goals conceded by team j in the last 2 competitive matches. 4. A home dummy. 5. A European Championship dummy to capture whether a team does systematically better at European Championships than in other competitive matches Second, we use these regression estimates and our assumed Poisson distribution in a Monte Carlo simulation with 100,000 draws to generate a distribution of outcomes for each of the 52 matches, from the opener between France and Romania on June 10 to the final on July 10. We use the rounded prediction of the goals scored to determine the outcome of each match during the group stage and the unrounded prediction to pick the winner in the knockout stage. Third, we use the estimation results to generate both a set of probabilities that a particular team reaches a particular stage of the tournament, up to and including the championship, and a modal—that is, single most likely—forecast for the outcome of each match, which we then run forward through the tournament until the final.


A SUMMARY OF PREDICTIONS
Our probabilities are shown in Exhibit 1. The model says that France has a 23% probability of winning the trophy, followed by Germany at 20%, Spain at 14%, and England at 11%. Although Germany has the highest Elo rating, France is favored because of its home advantage.




Exhibits 2 and 3 provide a different perspective by showing the modal prediction for the entire tournament. There are some interesting contrasts with the probabilities in Exhibit 1. For example, Exhibit 1 says that Germany is more likely than Spain to win the tournament because it is more likely to succeed across the entire range of possible tournament configurations. But Exhibit 3 says that in the single most likely case, Spain beats England in Semifinal 1, France beats Germany in Semifinal 2, and France then wins the final—i.e., Germany finishes behind Spain. Which approach is better, the probabilistic one in Exhibit 1 or the modal one in Exhibits 2 and 3? A modal forecast does have the advantage of being more “crisp.” The sentence “Goldman Sachs says France will win” has a better ring to it than “Goldman Sachs says France has a 23% probability of winning, with Germany close behind.” Nevertheless, we think that a probabilistic approach is more useful—for predicting the outcome of football tournaments and, increasingly, for our day-to-day work on economic forecasting.



Exhibit 4 provides more insight into the results by breaking down the probabilities of winning for the top four teams in a “waterfall chart” format. It shows that the most important factor is the Elo score, followed by home advantage and the European Championship dummy. The chart illustrates that the front-runner position for France derives largely from its home advantage, as its Elo rating is well below Germany’s and also a bit below Spain’s. Meanwhile, Germany benefits from the European Championship dummy, which picks up its historically strong tournament performance. 





HOW CONFIDENT CAN WE BE?

It is difficult to assess how much faith one should have in these predictions. On the plus side, our approach carefully considers the stochastic nature of the tournament using statistical methods, and we do think that the Elo rating—the most important input into our analysis—is a compelling summary of a team’s track record. On the minus side, we ignore a number of potentially important factors that are difficult to summarize statistically, including the quality of the individual players unless they are reflected in the team’s recent track record.[4] And there is no room for human judgment (which may not be such a bad thing given that none of us are really football experts but some are enthusiastic Germany supporters).[5]



One useful cross-check is to compare our results with bookmakers’ odds. Exhibit 5 plots our estimated championship probability against the average probability implied by the odds offered by five different bookmakers. The basic result is clear. Even though our model does not include bookmakers’ odds in any way, the probabilities are quite similar. A possible reason is that professional betting firms use many of the same inputs—such as Elo ratings—in their analysis and that they process the information in ways that are ultimately similar to ours. Another useful check is to evaluate the performance of our model for the 2014 World Cup, which followed an essentially identical approach to the one presented here.[6] It is safe to say that we had our hits and misses. First, performance in the group stage was not great. The model only identified 9 of the 16 advancing teams and failed to predict the elimination of heavyweights Spain and Italy, although it correctly anticipated that England would fly home early. Second, the model gave Brazil a 48% probability of winning the trophy, by far the highest of all the contestants. That failure illustrates a certain lack of imagination that is inherent in our approach. If the greatest football nation on earth—in terms of both past victories and its 2014 Elo rating—plays a World Cup at home, we are bound to project success. At least the probability was below 50%! Third, the model did correctly identify three of the four semifinalists before the start of the tournament, namely Argentina, Brazil, and Germany, although it incorrectly picked Spain over the Netherlands. Fourth, the fully updated version of the model—that is, the projection we sent out before each day of play on the basis of updated Elo ratings and other performance measures—was remarkably accurate during the knockout stage. It correctly predicted the winner of every match except the 7-1 semifinal between Germany and Brazil. But that was, by one estimate, the single most surprising result in World Cup history.[7] Ultimately, this last predictive failure might best capture the spirit of the exercise. As we said in our comment at the time: “Speaking as forecasters, we regret the miss. But, speaking as Germans, we would note that there are more important things than being right.” May the best team win and let’s hope that watching Euro 2016 is as much fun as it was to write this article!

10 Minutes to Data Science

I just followed John Hopkin's Executive Data Science team. In the first chapter of the course they talk,

*In Data Science, the importance is science and not data. Data Science is only useful when we use data to answer the question.*
It actually true in some point. I've actually seen too many companies that brag how big they data are but they don't actually know how to pose a question. If the data can't be used to answer the question that makes the company growth, it's not a good idea. They press even more to the point of data, data, data but really is in the end, machine learning will just feed into it, to make a better prediction to answer the question. It's critical to pose a question first, then try to get/build data to answer the question. And more importantly, don't be afraid to get other data from sources outside your company.
When investigating a problem and communicating to the broader audience, it's important to find the right question to answer, and then find the data related to answering the questions. This how all data science work will look like. Even in A/B testing when we measure the confidence interval between control and experiment group, the real question is still about how we can use data to answer our question.
In that course, Jeff Leek continues in term of Money Ball. We can find some evaluation metrics to measure player's skills, but the key important to answer questions is, "Can we be a winning team with a small budget?" Creating the best predictive model is not the most important, in the case of Netflix prize, one-million-dollar algorithm can't be implemented because it's impossible to scale with all of the customer combined.


Statistics and Machine Learning

Statistics mainly divided into two parts, Descriptive Statistics and Inferential Statistics. Descriptive Statistics, as the name implied, is using statistics to better understand the data. This includes using summary statistics and visualization to explore the data. Jake Vanderplas, the author of Python Data Science Handbook, showed how to use this method to understand the pattern of Seattle's Bicycle Habits in his blog.
While Inferential Statistics let you use Hypothesis testing and Confidence Interval to make an inference about your assumption. Suppose you have two group with numerical variable (female age vs male age), and you want to found whether these groups is significantly different with each other. Statistical Inference need you to follow experiment design, so that your inference can generalize well to your population of interest, and also found correlation that suggest causation. This method is useful to get insight, whether the relationship of two variable is correlate with each other.
Machine Learning is one of the fields in artificial intelligence that give a machine a capability to learn about your data. Thanks to the modern era, where computers have grown computation power and hype of data science , machine learning has grown into broad area. Two of the interesting topic is Supervised Learning and Unsupervised Learning. Supervised Learning is where given a set of input and output, machine learning tried to predict future output given future input. Think of it as a student that given a bunch of papers of quiz where it has a wrong and right answer, he learns and will able to answer the quiz. While the example of Unsupervised Learning is clustering algorithm like we discussed earlier.
Machine Learning is a different use case when compared to Statistical Inference. Have you seen Kaggle competitions? take a look around at their leaderboard, and all that score is extremely close. Often top 100 hundred is already a winner. They're willing to go into 2-3 times more complexity to get 1% increase. This is not practically possible, and we saw from Netflix prize, they give up on the one million dollar code because it's computation is expensive. So if you in for the accuracy games, go ahead! Otherwise, a simpler model is better.
So what do you use when you want to make a prediction? use Statistical modeling to understand what your prediction is. Use machine learning to make your prediction better. Statistical modeling concerns about the complexity model because we have to understand it, but machine learning scale as complexity increase. Moreover, Machine Learning concern about parameter tuning to performance, statistics concern about parsimonious model (get understanding bett

Software Engineering

So why is software engineering is important in data science? Because often you will have to get data using programming language. Sure there is some reporting that you can be download from Google Analytics or any dashboard in your company, but it's only summary and aggregation metrics. Things will get harder if you need advanced or specialized metrics that would need you to create aggregation yourself. Log data will always have some messy data. And if you have human-input data, there's going to be a human error. In order to do that, you need engineering skills. Pulling data from database alone need some programming or SQL skills. In case you need additional data from open data in the web, you also need programming to get data through API.
So engineering skill is a critical part of data science. In fact, it's so critical that you won't get the fun part, analyzing and making inference/prediction if you don't have engineering skills. You don't even know how to get data and clean it. At least software engineer alone can still do something. They can get some data and validate through analyzing inconsistency, and make some descriptive statistics to analyze the data.


Toolbox

One of data science toolbox is often a debate between choosing R or Python. But they are two different things. Python comes from software engineer background. Since Python famous first at web development, gathering and manipulating data has becomes very important. On the other hand, R has statistician background. There is widely variant of statistical packages available. And as statistician also use visualization to get insight about data, it also rich with visualization packages. So when processing the data use Python, and doing R when you want to analyze. Of course, sometimes these two overlaps, and you can choose to stick to one language. You can manipulate string in R, or do statistical analysis in Python. You have few choices when you want to present your result of analysis. If narrative, you can use Jupyter or Rmarkdown to storytelling your analysis in a narrative way. Sometimes you end-up want to engage your audience based on your findings. In this case, you want to create interactive visualization. D3.js is great in this area.
So when you talk about data science, think more about the question that you want to answer. Can you get some information if you have to answer the question? Do you have useful data to answer your question? If you could even answer it, is the answer practically possible? By doing this series of questions, it will avoid your missteps in the long run.

Python in one Picture

Predictive Analytics

Predictive analytics encompasses a variety of techniques from statistics, data mining and game theory that analyze current and historical facts to make predictions about future events.

In business, predictive models exploit patterns found in historical and transactional data to identify risks and opportunities. Models capture relationships among many factors to allow assessment of risk or potential associated with a particular set of conditions, guiding decision making for candidate transactions.

Predictive analytics is used in actuarial science, financial services, insurance, telecommunications, retail, travel, healthcare, pharmaceuticals and other fields.

One of the most well-known applications is credit scoring, which is used throughout financial services. Scoring models process a customer’s credit history, loan application, customer data, etc., in order to rank-order individuals by their likelihood of making future credit payments on time. A well-known example would be the FICO score.

Definition Predictive analytics is an area of statistical analysis that deals with extracting information from data and using it to predict future trends and behavior patterns. The core of predictive analytics relies on capturing relationships between explanatory variables and the predicted variables from past occurrences, and exploiting it to predict future outcomes. It is important to note, however, that the accuracy and usability of results will depend greatly on the level of data analysis and the quality of assumptions.

Types: Generally, the term predictive analytics is used to mean predictive modeling, "scoring" data with predictive models, and forecasting. However, people are increasingly using the term to describe related analytical disciplines, such as descriptive modeling and decision modeling or optimization. These disciplines also involve rigorous data analysis, and are widely used in business for segmentation and decision making, but have different purposes and the statistical techniques underlying them vary.

Predictive models: Predictive models analyze past performance to assess how likely a customer is to exhibit a specific behavior in the future in order to improve marketing effectiveness. This category also encompasses models that seek out subtle data patterns to answer questions about customer performance, such as fraud detection models. Predictive models often perform calculations during live transactions, for example, to evaluate the risk or opportunity of a given customer or transaction, in order to guide a decision. With advancement in computing speed, individual agent modeling systems can simulate human behavior or reaction to given stimuli or scenarios. The new term for animating data specifically linked to an individual in a simulated environment is avatar analytics.

Descriptive models: Descriptive models quantify relationships in data in a way that is often used to classify customers or prospects into groups. Unlike predictive models that focus on predicting a single customer behavior (such as credit risk), descriptive models identify many different relationships between customers or products. Descriptive models do not rank-order customers by their likelihood of taking a particular action the way predictive models do. Descriptive models can be used, for example, to categorize customers by their product preferences and life stage. Descriptive modeling tools can be utilized to develop further models that can simulate large number of individualized agents and make predictions.

Decision models: Decision models describe the relationship between all the elements of a decision — the known data (including results of predictive models), the decision and the forecast results of the decision — in order to predict the results of decisions involving many variables. These models can be used in optimization, maximizing certain outcomes while minimizing others. Decision models are generally used to develop decision logic or a set of business rules that will produce the desired action for every customer or circumstance.

Applications: Although predictive analytics can be put to use in many applications, we outline a few examples where predictive analytics has shown positive impact in recent years.

Analytical customer relationship management (CRM): Analytical Customer Relationship Management is a frequent commercial application of Predictive Analysis. Methods of predictive analysis are applied to customer data to pursue CRM objectives.

Clinical decision support systems: Experts use predictive analysis in health care primarily to determine which patients are at risk of developing certain conditions, like diabetes, asthma, heart disease and other lifetime illnesses. Additionally, sophisticated clinical decision support systems incorporate predictive analytics to support medical decision making at the point of care. A working definition has been proposed by Dr. Robert Hayward of the Centre for Health Evidence: "Clinical Decision Support systems link health observations with health knowledge to influence health choices by clinicians for improved health care."

Collection analytics: Every portfolio has a set of delinquent customers who do not make their payments on time. The financial institution has to undertake collection activities on these customers to recover the amounts due. A lot of collection resources are wasted on customers who are difficult or impossible to recover. Predictive analytics can help optimize the allocation of collection resources by identifying the most effective collection agencies, contact strategies, legal actions and other strategies to each customer, thus significantly increasing recovery at the same time reducing collection costs.

Cross-sell: Often corporate organizations collect and maintain abundant data (e.g. customer records, sale transactions) and exploiting hidden relationships in the data can provide a competitive advantage to the organization. For an organization that offers multiple products, an analysis of existing customer behavior can lead to efficient cross sell of products. This directly leads to higher profitability per customer and strengthening of the customer relationship. Predictive analytics can help analyze customers’ spending, usage and other behavior, and help cross-sell the right product at the right time.

Customer retention: With the number of competing services available, businesses need to focus efforts on maintaining continuous consumer satisfaction. In such a competitive scenario, consumer loyalty needs to be rewarded and customer attrition needs to be minimized. Businesses tend to respond to customer attrition on a reactive basis, acting only after the customer has initiated the process to terminate service. At this stage, the chance of changing the customer’s decision is almost impossible. Proper application of predictive analytics can lead to a more proactive retention strategy. By a frequent examination of a customer’s past service usage, service performance, spending and other behavior patterns, predictive models can determine the likelihood of a customer wanting to terminate service sometime in the near future. An intervention with lucrative offers can increase the chance of retaining the customer. Silent attrition is the behavior of a customer to slowly but steadily reduce usage and is another problem faced by many companies. Predictive analytics can also predict this behavior accurately and before it occurs, so that the company can take proper actions to increase customer activity.

Direct marketing: When marketing consumer products and services there is the challenge of keeping up with competing products and consumer behavior. Apart from identifying prospects, predictive analytics can also help to identify the most effective combination of product versions, marketing material, communication channels and timing that should be used to target a given consumer. The goal of predictive analytics is typically to lower the cost per order or cost per action.

Fraud detection: Fraud is a big problem for many businesses and can be of various types. Inaccurate credit applications, fraudulent transactions (both offline and online), identity thefts and false insurance claims are some examples of this problem. These problems plague firms all across the spectrum and some examples of likely victims are credit card issuers, insurance companies, retail merchants, manufacturers, business to business suppliers and even services providers. This is an area where a predictive model is often used to help weed out the “bads” and reduce a business's exposure to fraud.

Predictive modeling can also be used to detect financial statement fraud in companies, allowing auditors to gauge a company's relative risk, and to increase substantive audit procedures as needed.

The Internal Revenue Service (IRS) of the United States also uses predictive analytics to try to locate tax fraud.

Portfolio, product or economy level prediction: Often the focus of analysis is not the consumer but the product, portfolio, firm, industry or even the economy. For example a retailer might be interested in predicting store level demand for inventory management purposes. Or the Federal Reserve Board might be interested in predicting the unemployment rate for the next year. These type of problems can be addressed by predictive analytics using Time Series techniques (see below).

Underwriting: Many businesses have to account for risk exposure due to their different services and determine the cost needed to cover the risk. For example, auto insurance providers need to accurately determine the amount of premium to charge to cover each automobile and driver. A financial company needs to assess a borrower’s potential and ability to pay before granting a loan. For a health insurance provider, predictive analytics can analyze a few years of past medical claims data, as well as lab, pharmacy and other records where available, to predict how expensive an enrollee is likely to be in the future. Predictive analytics can help underwriting of these quantities by predicting the chances of illness, default, bankruptcy, etc. Predictive analytics can streamline the process of customer acquisition, by predicting the future risk behavior of a customer using application level data. Predictive analytics in the form of credit scores have reduced the amount of time it takes for loan approvals, especially in the mortgage market where lending decisions are now made in a matter of hours rather than days or even weeks. Proper predictive analytics can lead to proper pricing decisions, which can help mitigate future risk of default.

Statistical techniques: The approaches and techniques used to conduct predictive analytics can broadly be grouped into regression techniques and machine learning techniques.

Regression techniques: Regression models are the mainstay of predictive analytics. The focus lies on establishing a mathematical equation as a model to represent the interactions between the different variables in consideration. Depending on the situation, there is a wide variety of models that can be applied while performing predictive analytics. Some of them are briefly discussed below.

Linear regression model: The linear regression model analyzes the relationship between the response or dependent variable and a set of independent or predictor variables. This relationship is expressed as an equation that predicts the response variable as a linear function of the parameters. These parameters are adjusted so that a measure of fit is optimized. Much of the effort in model fitting is focused on minimizing the size of the residual, as well as ensuring that it is randomly distributed with respect to the model predictions.

The goal of regression is to select the parameters of the model so as to minimize the sum of the squared residuals. This is referred to as ordinary least squares (OLS) estimation and results in best linear unbiased estimates (BLUE) of the parameters if and only if the Gauss-Markov assumptions are satisfied.

Once the model has been estimated we would be interested to know if the predictor variables belong in the model – i.e. is the estimate of each variable’s contribution reliable? To do this we can check the statistical significance of the model’s coefficients which can be measured using the t-statistic. This amounts to testing whether the coefficient is significantly different from zero. How well the model predicts the dependent variable based on the value of the independent variables can be assessed by using the R² statistic. It measures predictive power of the model i.e. the proportion of the total variation in the dependent variable that is “explained” (accounted for) by variation in the independent variables.

Discrete choice models: Multivariate regression (above) is generally used when the response variable is continuous and has an unbounded range. Often the response variable may not be continuous but rather discrete. While mathematically it is feasible to apply multivariate regression to discrete ordered dependent variables, some of the assumptions behind the theory of multivariate linear regression no longer hold, and there are other techniques such as discrete choice models which are better suited for this type of analysis. If the dependent variable is discrete, some of those superior methods are logistic regression, multinomial logit and probit models. Logistic regression and probit models are used when the dependent variable is binary.

Logistic regression: For more details on this topic, see logistic regression.
In a classification setting, assigning outcome probabilities to observations can be achieved through the use of a logistic model, which is basically a method which transforms information about the binary dependent variable into an unbounded continuous variable and estimates a regular multivariate model (See Allison’s Logistic Regression for more information on the theory of Logistic Regression).

The Wald and likelihood-ratio test are used to test the statistical significance of each coefficient b in the model (analogous to the t tests used in OLS regression; see above). A test assessing the goodness-of-fit of a classification model is the –.

Multinomial logistic regression: An extension of the binary logit model to cases where the dependent variable has more than 2 categories is the multinomial logit model. In such cases collapsing the data into two categories might not make good sense or may lead to loss in the richness of the data. The multinomial logit model is the appropriate technique in these cases, especially when the dependent variable categories are not ordered (for examples colors like red, blue, green). Some authors have extended multinomial regression to include feature selection/importance methods such as Random multinomial logit.

Probit regressionProbit models offer an alternative to logistic regression for modeling categorical dependent variables. Even though the outcomes tend to be similar, the underlying distributions are different. Probit models are popular in social sciences like economics.

A good way to understand the key difference between probit and logit models, is to assume that there is a latent variable z.

We do not observe z but instead observe y which takes the value 0 or 1. In the logit model we assume that y follows a logistic distribution. In the probit model we assume that y follows a standard normal distribution. Note that in social sciences (example economics), probit is often used to model situations where the observed variable y is continuous but takes values between 0 and 1.

Logit versus probit: The Probit model has been around longer than the logit model. They look identical, except that the logistic distribution tends to be a little flat tailed. One of the reasons the logit model was formulated was that the probit model was difficult to compute because it involved calculating difficult integrals. Modern computing however has made this computation fairly simple. The coefficients obtained from the logit and probit model are also fairly close. However, the odds ratio makes the logit model easier to interpret.

For practical purposes the only reasons for choosing the probit model over the logistic model would be:

There is a strong belief that the underlying distribution is normal
The actual event is not a binary outcome (e.g. Bankrupt/not bankrupt) but a proportion (e.g. Proportion of population at different debt levels).
Time series models: Time series models are used for predicting or forecasting the future behavior of variables. These models account for the fact that data points taken over time may have an internal structure (such as autocorrelation, trend or seasonal variation) that should be accounted for. As a result standard regression techniques cannot be applied to time series data and methodology has been developed to decompose the trend, seasonal and cyclical component of the series. Modeling the dynamic path of a variable can improve forecasts since the predictable component of the series can be projected into the future.

Time series models estimate difference equations containing stochastic components. Two commonly used forms of these models are autoregressive models (AR) and moving average (MA) models. The Box-Jenkins methodology (1976) developed by George Box and G.M. Jenkins combines the AR and MA models to produce the ARMA (autoregressive moving average) model which is the cornerstone of stationary time series analysis. ARIMA (autoregressive integrated moving average models) on the other hand are used to describe non-stationary time series. Box and Jenkins suggest differencing a non stationary time series to obtain a stationary series to which an ARMA model can be applied. Non stationary time series have a pronounced trend and do not have a constant long-run mean or variance.

Box and Jenkins proposed a three stage methodology which includes: model identification, estimation and validation. The identification stage involves identifying if the series is stationary or not and the presence of seasonality by examining plots of the series, autocorrelation and partial autocorrelation functions. In the estimation stage, models are estimated using non-linear time series or maximum likelihood estimation procedures. Finally the validation stage involves diagnostic checking such as plotting the residuals to detect outliers and evidence of model fit.

In recent years time series models have become more sophisticated and attempt to model conditional heteroskedasticity with models such as ARCH (autoregressive conditional heteroskedasticity) and GARCH (generalized autoregressive conditional heteroskedasticity) models frequently used for financial time series. In addition time series models are also used to understand inter-relationships among economic variables represented by systems of equations using VAR (vector autoregression) and structural VAR models.

Survival or duration analysis: Survival analysis is another name for time to event analysis. These techniques were primarily developed in the medical and biological sciences, but they are also widely used in the social sciences like economics, as well as in engineering (reliability and failure time analysis).

Censoring and non-normality, which are characteristic of survival data, generate difficulty when trying to analyze the data using conventional statistical models such as multiple linear regression. The normal distribution, being a symmetric distribution, takes positive as well as negative values, but duration by its very nature cannot be negative and therefore normality cannot be assumed when dealing with duration/survival data. Hence the normality assumption of regression models is violated.

The assumption is that if the data were not censored it would be representative of the population of interest. In survival analysis, censored observations arise whenever the dependent variable of interest represents the time to a terminal event, and the duration of the study is limited in time.

An important concept in survival analysis is the hazard rate, defined as the probability that the event will occur at time t conditional on surviving until time t. Another concept related to the hazard rate is the survival function which can be defined as the probability of surviving to time t.

Most models try to model the hazard rate by choosing the underlying distribution depending on the shape of the hazard function. A distribution whose hazard function slopes upward is said to have positive duration dependence, a decreasing hazard shows negative duration dependence whereas constant hazard is a process with no memory usually characterized by the exponential distribution. Some of the distributional choices in survival models are: F, gamma, Weibull, log normal, inverse normal, exponential etc. All these distributions are for a non-negative random variable.

Duration models can be parametric, non-parametric or semi-parametric. Some of the models commonly used are Kaplan-Meier and Cox proportional hazard model (non parametric).

Classification and regression trees: Main article: decision tree learning
Classification and regression trees (CART) is a non-parametric decision tree learning technique that produces either classification or regression trees, depending on whether the dependent variable is categorical or numeric, respectively.

Decision trees are formed by a collection of rules based on variables in the modeling data set:

Rules based on variables’ values are selected to get the best split to differentiate observations based on the dependent variable
Once a rule is selected and splits a node into two, the same process is applied to each “child” node (i.e. it is a recursive procedure)
Splitting stops when CART detects no further gain can be made, or some pre-set stopping rules are met. (Alternatively, the data is split as much as possible and then the tree is later pruned.)
Each branch of the tree ends in a terminal node. Each observation falls into one and exactly one terminal node, and each terminal node is uniquely defined by a set of rules.

A very popular method for predictive analytics is Leo Breiman's Random forests or derived versions of this technique like Random multinomial logit.

Multivariate adaptive regression splines: Multivariate adaptive regression splines (MARS) is a non-parametric technique that builds flexible models by fitting piecewise linear regressions.

An important concept associated with regression splines is that of a knot. Knot is where one local regression model gives way to another and thus is the point of intersection between two splines.

In multivariate and adaptive regression splines, basis functions are the tool used for generalizing the search for knots. Basis functions are a set of functions used to represent the information contained in one or more variables. Multivariate and Adaptive Regression Splines model almost always creates the basis functions in pairs.

Multivariate and adaptive regression spline approach deliberately overfits the model and then prunes to get to the optimal model. The algorithm is computationally very intensive and in practice we are required to specify an upper limit on the number of basis functions.

Machine learning techniques. Machine learning, a branch of artificial intelligence, was originally employed to develop techniques to enable computers to learn. Today, since it includes a number of advanced statistical methods for regression and classification, it finds application in a wide variety of fields including medical diagnostics, credit card fraud detection, face and speech recognition and analysis of the stock market. In certain applications it is sufficient to directly predict the dependent variable without focusing on the underlying relationships between variables. In other cases, the underlying relationships can be very complex and the mathematical form of the dependencies unknown. For such cases, machine learning techniques emulate human cognition and learn from training examples to predict future events.

A brief discussion of some of these methods used commonly for predictive analytics is provided below. A detailed study of machine learning can be found in Mitchell (1997).

Neural networks. Neural networks are nonlinear sophisticated modeling techniques that are able to model complex functions. They can be applied to problems of prediction, classification or control in a wide spectrum of fields such as finance, cognitive psychology/neuroscience, medicine, engineering, and physics.

Neural networks are used when the exact nature of the relationship between inputs and output is not known. A key feature of neural networks is that they learn the relationship between inputs and output through training. There are two types of training in neural networks used by different networks, supervised and unsupervised training, with supervised being the most common one.

Some examples of neural network training techniques are backpropagation, quick propagation, conjugate gradient descent, projection operator, Delta-Bar-Delta etc. Some unsupervised network architectures are multilayer perceptrons, Kohonen networks, Hopfield networks, etc.

Radial basis functions: A radial basis function (RBF) is a function which has built into it a distance criterion with respect to a center. Such functions can be used very efficiently for interpolation and for smoothing of data. Radial basis functions have been applied in the area of neural networks where they are used as a replacement for the sigmoidal transfer function. Such networks have 3 layers, the input layer, the hidden layer with the RBF non-linearity and a linear output layer. The most popular choice for the non-linearity is the Gaussian. RBF networks have the advantage of not being locked into local minima as do the feed-forward networks such as the multilayer perceptron.

Support vector machines: Support Vector Machines (SVM) are used to detect and exploit complex patterns in data by clustering, classifying and ranking the data. They are learning machines that are used to perform binary classifications and regression estimations. They commonly use kernel based methods to apply linear classification techniques to non-linear classification problems. There are a number of types of SVM such as linear, polynomial, sigmoid etc.

Naïve Bayes: Naïve Bayes based on Bayes conditional probability rule is used for performing classification tasks. Naïve Bayes assumes the predictors are statistically independent which makes it an effective classification tool that is easy to interpret. It is best employed when faced with the problem of ‘curse of dimensionality’ i.e. when the number of predictors is very high.

K-nearest neighbours: The nearest neighbour algorithm (KNN) belongs to the class of pattern recognition statistical methods. The method does not impose a priori any assumptions about the distribution from which the modeling sample is drawn. It involves a training set with both positive and negative values. A new sample is classified by calculating the distance to the nearest neighbouring training case. The sign of that point will determine the classification of the sample. In the k-nearest neighbour classifier, the k nearest points are considered and the sign of the majority is used to classify the sample. The performance of the kNN algorithm is influenced by three main factors: (1) the distance measure used to locate the nearest neighbours; (2) the decision rule used to derive a classification from the k-nearest neighbours; and (3) the number of neighbours used to classify the new sample. It can be proved that, unlike other methods, this method is universally asymptotically convergent, i.e.: as the size of the training set increases, if the observations are independent and identically distributed (i.i.d.), regardless of the distribution from which the sample is drawn, the predicted class will converge to the class assignment that minimizes misclassification error. 
Geospatial predictive modeling: Conceptually, geospatial predictive modeling is rooted in the principle that the occurrences of events being modeled are limited in distribution. Occurrences of events are neither uniform nor random in distribution – there are spatial environment factors (infrastructure, sociocultural, topographic, etc.) that constrain and influence where the locations of events occur. Geospatial predictive modeling attempts to describe those constraints and influences by spatially correlating occurrences of historical geospatial locations with environmental factors that represent those constraints and influences. Geospatial predictive modeling is a process for analyzing events through a geographic filter in order to make statements of likelihood for event occurrence or emergence.

Tools: There are numerous tools available in the marketplace which help with the execution of predictive analytics. These range from those which need very little user sophistication to those that are designed for the expert practitioner. The difference between these tools is often in the level of customization and heavy data lifting allowed.

In an attempt to provide a standard language for expressing predictive models, the Predictive Model Markup Language (PMML) has been proposed. Such an XML-based language provides a way for the different tools to define predictive models and to share these between PMML compliant applications. PMML 4.0 was released in June, 2009.