Throughout these analyzes, we will refer to a model we use to analyze clients. It may be useful to scroll down to the very last section to read an overview of the model.

Big trends

A big trend in this cycle was the outflow of traditional US-targeting clients going after foreign offices. This was a huge shift from last year. Fully, 67% of our successful clients placed in foreign offices. In fact, clients declined by a U.S. office were applying to Johannesburg, Dubai, Korea, Singapore, etc. and securing interviews, and ultimately offers. We are not saying foreign offices have necessarily an easier interview process, but the supply of quality candidates is lower. The PhD’s were a much larger group here.

The number of females was sharply higher in our Sep/Oct client group.

• Of the successful placements, 53% were female.

• Unlike male PhD’s and male MBA clients, females were predominantly from the traditionally strong placement schools like Yale, Princeton, Harvard, Wharton, Stanford etc.

• 76% of the 53% successful female placements originated from these schools.

• That said, the majority possessed undergraduate degrees from largely unknown schools, to us, or foreign schools.

Foreign students studying in the US dominated this group.

PhD performance

We fully expected and prepared for a weaker showing in the PhD group. We were wrong of course, but spent a long time analyzing the numbers and figuring out why our PhD clients did so much better when we expected worse results.

We have a good hypothesis.

Our belief is that because we expected PhD clients to fare worse, we dramatically over compensated for this expectation, and this over compensation gave them an advantage.

For many PhD clients we hoped for the best, but planned for the worst. We expected weaker results and were counselling clients well into their final rounds that if things did not work out, there was always next year, or if a different office was an option, then just a 4-6 month delay if they knew how to navigate the system.

Our belief is that because we expected PhD clients to fare worse, we dramatically over compensated for this expectation, and this over compensation gave them an advantage.

This is an important point worth explaining. The U.S.A. churns out PhDs like a McDonalds restaurant’s first week of opening in an emerging market. Knowing this, we were significantly tougher on PhD’s writing resumes, cover letters and networking. We treated master’s students – non-MBA – and experienced hires in the same way. So when we say PhDs, we refer to this group as well.

Given the long networking lead time, we insisted most PhD clients create a 3-6 month preparation window with us.

Let’s explain this in another way. We felt PhD’s would do worse and gave them more work and lead times. When we inputted these variables into the model, we expected the model to say, “Yes”, they will do worse. However, the model says, based on historical patterns, stripping out emotion, the way we prepared PhD candidates means they would do better, not worse.

In hindsight this makes perfect sense. Since the model looks at past performance, it looks at a current client’s attributes and assigns them to a bucket of performance. The model does not know longer lead times will lead to a poorer performance unless we tell the model that it does. And of course, the “telling the model” part comes from past data which told it something else entirely.

The other surprise among PhD candidates was the strong showing of California. We normally do not expect California-based clients to dominate the final results. They did. 22% of successful PhD candidates were from California schools.

We personally do not see this as the rise of any underlying trend. Given the way we work, heavily referral-based, we typically expect a spike in a region a few months after a few successful placements the year before – clients unfortunately talk to others.

That said, the sheer number of PhD candidates on the East Coast and the increase in hiring, will mean that all other things being equal, California-based client placements as a percentage of the total will drop next year.

Great results without exceptional resumes

We have Rhodes Scholars, Marshall Scholars, and Fulbright Scholars etc. in our program. None of them were in the group which placed well. Every one of that group had their application dates or interviews pushed back for a variety of reasons.

That means the group which placed highly did not have extraordinary profiles. They were good, but not exceptionally so on paper. This is an important observation which the model supports and which we can concur through our placement and general client observations.

Clients with extraordinary backgrounds, like Rhodes scholars etc., fit into a binary pattern. They either come across extraordinarily well or extraordinarily poorly. There is very little “average” performance. This is a consistent trait we find.

Therefore when screening candidates with superb paper credentials we don’t worry about the positive outliers, unless they are arrogant, which will also not go down well with partners, but more about those who either struggle to mirror their paper profile or struggle to communicate.

The communication hurdle is a huge problem and much harder to overcome.

Differences between candidates on their second attempt at MBB

Between 37% and 56% of our clients were on their second attempt at McKinsey et al, though their first time working with us. The range exists purely based on how you define the first attempt: failed PST, failed first round, failed final round etc. We do notice that candidates, who already have their first strike, if you really want to keep the California theme going, tend to be more disciplined and “mature” about the process.

You tend to have a uniformly bleak view of the world when you are 32, sitting in a lab, earning $50K/annum or less, and you just have one shot with McKinsey.

There is a deeper and profound sense of urgency in most cases. I am not saying it is uniform, but someone who believes they can reapply in the future tends to assume failure is just temporary and should be shrugged off. This is supported by both the model and our own observations.

Again, this theme is far more pronounced among PhD’s and experienced hires versus MBA’s.

Causality is far easier to explain here.

A second-attempt for a PhD is going to be for someone who is 27, at the minimum, and possibly 38, at the upper limit. The majority sit at around 32 years which means there is far more at stake.

You tend to have a uniformly bleak view of the world when you are 32, sitting in a lab, earning $50K/annum or less, and you just have one shot with McKinsey.

The urgency arrives.

The same applies if you are in industry, have a family and have reached a career ceiling.

Negative trends in candidates

However, too much urgency quickly devolves into 4 trends which both the model and our own experience quickly validate.

Candidates who think they have only one chance or a small chance tend to believe there is some magical advice ricocheting through the corridors of their school or the dark alleys of their friendships.

The first negative trend is the inability to deduce what is important advice, and by default, creates the tendency to follow any and all advice. Of the candidates who do not get offers, about 50%, in our opinion, fall into this trap.

The model predicts a full 72% fell into trap. I would say here the model is probably more accurate since this is a common problem. Candidates who think they have only one chance or a small chance tend to believe there is some magical advice ricocheting through the corridors of their school or the dark alleys of their friendships.

They make every effort to find this advice, which is a tiring and confusing process of networking as heavily as possible, not filtering the advice received and then trying to follow all the advice even when it conflicting.

A common example of this is the MBA student who feels it is their patriotic duty, no, their family obligation, to partake in every mock interview offered in the day. Our advice on this is clear. If you have not been trained yet, you are not practicing; you are simply learning, and learning from people who should not be teaching – the blind following the blind.

Candidates end up tired – it is draining to do 3 cases with unprepared people – and confused since we have not explained even 1/10 of what they faced in those poorly managed sessions. There are many other examples of this, but if all advice is treated like special advice then no advice is truly special.

The second negative trend is too much pandering in the system. When a client refers to any of our coaches as “sir” or “madam” that sets off an immediate warning bell in the system.

This is a simple one and the model also tends to be quite accurate here. Consulting firms are not looking for people who do not have the confidence to build relationships at a peer level. When clients act like this with our partners, they tend to replicate this behaviour with interviewers.

Understandably, culture may be the reason. While that reason is valid, and should be respected, it does not change the outcome.

That means clients who lack confidence put themselves into worse positions, which breed greater desperation, generating less confidence and leading to more desperate behaviour which only worsens the situation.

This leads to the third trend. Confidence is seen through every little thing you do or say. Our defining rule is never to look desperate. There is never an appropriate time when a strategy of desperation will work.

We can preach this as much as possible, but a full 90% of clients who do not make it exhibit this characteristic in one form or another. And here is a tip, being confident in 99 of your actions in an interview and showing lack of confidence in just 1, is usually enough, since it shows inconsistency.

The problem with confidence is that it is a cumulative degradation.

That means clients who lack confidence put themselves into worse positions, which breed greater desperation, generating less confidence and leading to more desperate behaviour which only worsens the situation.

The classic example: emailing a partner and asking if you are a fit for the firm and attaching your resume. 99% of clients seem pleased when the partner responds. You should not be. Asking if you deserve to be there implies you do not think you deserve to be there.

If you want to generate a conversation, do it in a way which does not sabotage your later plans.

The fourth trend here is a stunningly high number of people in the program, about 89% of those who do not make it seem to think that you need the poise and theatrical skills of a “movie star.” Consulting partners do not speak like this. We are not theatrical. We are analytical and professional.

If you want to work on communication focus on saying what you mean and meaning what you say: getting that part right is more important than merely sounding right.

New joiners to McKinsey et al need to show superior analytical skills and the ability to communicate in a simple manner. That means knowing what you want to say, and then figuring out how to say it.

Actors and actresses get told what to say and then do about 20 takes to get it right. The entire conversation is staged and scripted. You will not be like them and I dare you to find a partner who speaks like that.

Sounding good is not the same as being good. You need to work far more heavily on coming across naturally, but emphasizing strengths, versus being someone totally new.

The time needed to prepare

The model shows that clients who take 3 months or more to prepare have a 57% higher probability of getting an offer. Anecdotally, that makes perfect sense. Looking at the strong showing of PhD candidates, the vast majority were working with us between 3 to 9 months. Just one worked with us for 4 weeks, though she did get an offer at BCG.

Candidates who start early with us can regroup when things go wrong – and things go wrong often.The planned interview does not materialize, the PST is a hurdle too much, an office is not hiring, or the candidate becomes ill, gets married or has a kid. The more time you have, the easier it is to side-step these obstacles without taking desperate measures like emailing a partner and asking if you are a good fit etc.

The causality is very easy to outline here as well. Our strategy is always to create multiple paths for candidates. We call this having “options.” In our view, a candidate with just one or a few options is begging for disappointment. There are far too many variables outside our control to rely on a single option/path to an offer.

Candidates who start early with us can regroup when things go wrong – and things go wrong often.

The planned interview does not materialize, the PST is a hurdle too much, an office is not hiring, or the candidate becomes ill, gets married or has a kid. The more time you have, the easier it is to side-step these obstacles without taking desperate measures like emailing a partner and asking if you are a good fit etc.

Linked to this, candidates have more time to review lesson plans, listen to recordings, practice and share ideas with us. None of this is possible when the preparation timeline is short. We were personally happy when so many MBA clients signed up 6 months before interviews. That joy quickly dissipated when they only made themselves available 2 months before the interview, or did not allocate sufficient quality time when they did make themselves available.

You cannot predict the length of the figurative runway you need to prepare. Expect it to always be longer than needed.

The importance of rewriting the resume

The next lead indicator is linked to preparing early but still important to mention. The model shows those who made large changes to their resume, and made material improvements did better. However, the causal relationship is not what most would think. The improved resume helped, in some cases, it changed a smirk from the recruiter to “you must apply.”

Clients jumping straight into the coaching really struggle to understand the expectations we have of them, and the intensity of the program. The resume rewriting process leaves no doubts about the level of detail expected and the way we tackle things.

However, the main benefit is because it allowed us to get to know a candidate very well before the coaching began and this helped them know what to expect when the sessions actually began.

Clients jumping straight into the coaching really struggle to understand the expectations we have of them, and the intensity of the program. The resume rewriting process leaves no doubts about the level of detail expected and the way we tackle things.

The model again shows this, and I believe this is one area where the model tends to be quite accurate. Though, the curve is far from normal. There are outliers, who are confident, and do well immediately in the coaching, but they are a minority.

Again a related lead indicator which the model shows to be heavily correlated to successful placements has to do with flexibility from the client with regard to coaching dates and times. This one is again easy to explain. Clients who start early have flexibility and those rushing do not. It is that simple. I don’t think we needed a model to tell us that, but we needed to know the relationship to model its impact on the final result.

Too many practice sessions do not help

Here is the most interesting finding in the model. The probability of getting an offer is inversely correlated to the number of coaching sessions and practice sessions done.

Anecdotally, the causal relationship is again quite easy to understand. A candidate, who does not grasp the core consulting principles from the first six lessons, is then entering the next six sessions with a weak foundation. It is like building a huge mansion on a sink hole.

We find that candidates who have done more than 40-70 practice cases before we have trained them have reached a tipping point because they have learned too many bad habits. This is a statistically significant relationship.

Anecdotally, the causal relationship is again quite easy to understand. A candidate, who does not grasp the core consulting principles from the first six lessons, is then entering the next six sessions with a weak foundation. It is like building a huge mansion on a sink hole.

It’s not going to play out as we intended. Moreover, a client unable to understand the approach from the 12 sessions and the library of training videos is usually performing poorly due to weak study habits and basically inattentiveness.

Doing more sessions does not fix this problem. In fact, it makes the problem worse. It comes down to moral hazard. A client who believes they have a large supply of lessons or practice sessions will usually not develop the appropriate skills to extract all the lessons from the sessions they do have. We deliberately insist clients begin practicing soon and wean many off lessons.

No one likes it but it can only help them in the long term. This relationship was initially surprising from the model but the evidence clearly supports it. That said, some clients do benefit from longer sessions, provided they prepare well. They are easily the minority.

Wistia (the application hosting our files) generates a tonne of data for us on video usage. Again, the model and reality intersect well here. Clients who have access to the videos and do not use them well, almost always do not place well. This result in the model is almost perfectly correlated with reality.

However, this flag is not cast in stone. The videos are well explained, can be replayed countless times and clients who do not take the time or do not have the ability to understand them will fare poorly. In the latter case, they either just not attentive enough to put in the time or are unwilling to understand and use this competitive advantage.

Whenever a client enters a session unprepared we always try to understand why they were unprepared. The answer is almost always in the same areas: a) They watched the video a long time ago; b) They did not properly watch the videos which mean they typically skimmed through key parts or just looked at the structures, or c) they did not take the time to research and understand key ideas.

When a client enters a session without preparing appropriately, it indicates a problem outside of competence. It is a question of their focus, attitude and dedication. This is one of the largest red flags in the system.

This group who do not use the videos well can be broken down even further. Some clients are not sure how to regroup after a few weak starts. They end up doing badly. Another sub-group struggles, but takes the feedback and regroups well. A lot of PhDs fall into this group, even those who placed well. Therefore, poor use of the videos is not a bad sign if you can take feedback, go back and prepare. It requires more work and is not an easy route, but needs to be done.

Corporate finance clients

That said, one group is consistently strong on videos and their placement rates. Corporate finance clients have the highest video engagement numbers of any group and the single highest placement rate we have.

Though, we should caution this is a smaller group and this will ultimately normalize over time.

Moreover, this group has the highest referral score of all clients and corporate finance tends to be very technical so video guidance is required. That said, someone not understanding business would have the same relative deficit on general consulting issues and would need the general videos as well. Therefore, we still think it comes down to the attitude of this group. Fully 84% of corporate finance clients are referred which probably leads to like referring like. Corporate finance clients do however have the highest GPA’s of any client group, as well.

Over time we will refine the model and constantly update the data we extract from clients. As the relationships become clearer or sub-relationships are found, we hope to model that as well. That said, the model is not expected to replace human judgement but help us get a general feel of the bigger picture. Even where the model and our experience correlate perfectly, we still do not use the model output to draw unnecessary conclusions. It is simply an early warning device.

Understanding the model we built with client data:

Clients who worked with us know we ask for a lot of information throughout the interaction. We ask that every interaction be shared with us, and preferably a copy of any networking emails sent, be mailed to us as well. Basically, we want to know as much as the candidate.

We have now assembled all the data into one statistical model.

We track a significant amount of client data in this model, which allows us to run some very interesting statistical simulations.

We first track the easy-to-measure variables like:

• age,

• school,

• grades,

• GMAT,

• nationality,

• ethnicity,

• offices pursued,

• interviews secured,

• companies employed at,

• number of coaching sessions,

• performance in coaching sessions,

• length of coaching program,

• time before interviews,

• number of resume rewrites,

• number of cover letter rewrites,

• quality of rewrites,

• offices declined,

• types of emails received from offices,

• number of networking sessions,

• types of networking sessions,

• length of networking sessions,

• topic of networking sessions,

• level networked at,

• % leading to referrals,

• referrals leading to interviews, and

• the time before interviews etc.

We also track many smaller, but more important variables. All our coaching sessions are recorded and transcribed to text using Wistia. In many cases, clients also load practice sessions to Wistia. This is also transcribed to text. Wistia has some very interesting capabilities. Therefore, from the transcripts of each session can we can also electronically measure:

• the time to answer each question,

• number of mistakes made,

• type of mistakes made,

• language used,

• preparation of candidate,

• tone and energy of the call,

• number of questions asked,

• type of questions asked,

• length of time spent on training videos in the library,

• which sections of the videos were watched

• and how often parts of a video were skipped,

• podcasts listened to,

• how many times each podcast was listened to,

• Parts of podcasts skipped.

Finally we track administrative items, albeit subjectively, which provide vital clues to a candidate’s motivation:

• number of sessions a week,

• location of sessions,

• flexibility of sessions,

• decisiveness of decision making,

• comfort with decisions made,

• ability to take notes,

• quality of notes,

• ability to take and process feedback,

• number of additional sessions requested,

• tone in which requests are made,

• time to respond to emails,

• precision of communication,

• time set aside for practicing,

• quality of feedback from self-practicing etc.

We track many more variables. The point is that we now have what we consider to be a very interesting model to extract correlations between successful and less successful students. We admit it is far from perfect, but it is the most detailed database we have ever seen since we know each candidate personally and can verify all the data provided.

That is the most vital point, getting quality data.

Moreover, since all the recordings remain in our system, and we continue to track a candidate’s movement through the system well after the official coaching ends, the quality of the database can only improve as it collects more data.

We have looked at key variables which tend to have the greatest impact on obtaining an offer and tried to understand the nature of the relationship. We admit this is not an exact science and we constantly tweak the list of variables and the nature of the relationship. We have to go through substantial work to normalize the data, remove outliers etc.

Basically, it is the typical process to set up any social-science experiment.

That said, once we think the nature of the relationship is representative of the data, we can model a distribution curve for each variable, and they are not normal distributions at all. We then set limits to the distribution curves and determine the correlations between each distribution curve and the base variable.

We have determined the base variable to be a weighted average of grades and what we call a confidence index which we generate from data in the screening calls, and several other factors described below. So everything in the model is correlated back to these base variables.

We use a standard Monte Carlo “random” generator. As an aside, random is in inverted commas because the system is far from random. The point is that if we made the system random, the output would be spurious – statistical speak for wrong.

We remove the randomness by not using normal distributions/curves and also different limits for each curve we choose.

In simple terms the model works as follows. Let’s assume the base variables for one client is a GMAT score of 720, GPA of 3.80, from Princeton in economics, confidence index of 6/10 and just 3 weeks to go before interviews. The system then says, “Okay” the candidate now has 3 options, for example. They can prepare their cover letter and resume, ignore that step and move straight to an application or defer their application. Based on the candidates profile and past history of similar candidates in the database, it assigns a probability to each option and runs the simulation about 10,000 times. Overall, every possible path, even the lowest probability path, is modeled.

Assuming the model picked “moving straight to application” as the most probable option, it then lists the candidate’s exhaustive list of options at this stage. There can be anywhere from 3 to 8 mutually exclusive and collectively exhaustive options. Some steps can have 15 options. It repeats the process of assigning probabilities and choosing an option. Eventually, the model works its way down this chain of options, a very large decision-tree if you may, and arrives at a list of likely outcomes and their probabilities.

Obviously, as we move down each stage of the process with a candidate, we can compare their actual results with the simulated results. The model works well for clients who tend towards a mean, the average candidate, but less so for outliers. This is expected. We need to use judgment to manage the outliers, not assume they do not exist.

By now, you can probably figure out it is a rather large model consisting of 251 data sets – client files – of which 187 are complete, since we started collecting detailed data much later and we have to make some assumptions on the early clients for whom we do not have all data. For each data set, client, there are 233 variables measured with the majority collected in the coaching session and via Wistia which analyses the coaching session transcripts for patterns and trends.

We used this model to test different phenotypes – observable characteristics or variables if you want to use that word – of a client. So each time a candidate applies to join or a client joins the program, we would run the model. The model was only completed in late August so we used it less them we would have liked to. And as a model, which is not perfect, we could not rely on it too much. The key was to see if the model output matched reality.

So we ran the distribution curves we found and have attempted to explain the causality. Where the correlation existed and causality was missing, or we just could not rationally explain it, we left it out since it was interesting but not necessarily insightful.

In summary, the model predicted an unusual result in late August. It predicted an unusually high placement rate among PhDs. The highest we have ever had.

We eventually did have a very high placement rate among PhDs. That does not mean the model is correct. For all we know, it is possible that since we built the model, and obviously know each client personally, we could accurately judge their performance and expected them to do well, thereby building in correlations which reflected this reality. The test will come if the model, without any major tweaks, is run for a new group of clients whom we have not yet coached, and then fairly accurately predicts their performance.

Personally, we would be surprised if any model could do that. There are just too many random variables at play which we could never know and compensate for. So, my hypothesis is there must be some measure of bias which we cannot compensate for, because we cannot understand the complete nature of the bias.

For the stats majors, I am not going to get into a discussion on auto-correlations, stripping out seasonality, adjusting @Risk, the hundreds of assumptions documented or even the challenges of listing every single permutation and combination of options at each step.

That said the results are still very interesting and very useful for us to map out and plan a strategy for each client before they begin with us. Just knowing all their options and the probabilities gives us a major advantage. It allows us to institutionalize all the learning’s from working with previous clients.

For example, as coaches, we always knew that clients who don’t watch the videos or watch them sporadically do much worse than those who do. Yet, keeping track of such relationships in our heads is not a good way to use that knowledge, nor is it sustainable.

It is just not possible to login and check all this information – over 200 variables – for each client before each session. That would be an inefficient use of time even if it were possible. And if we did it, how could we know what the impact of the relationships is?

Aggregating all this data means we can rely on our historical lessons to a greater extent and, hopefully, not repeat as many mistakes.

That can only be an improvement.