Patrick Ng - Drill Down Into the Risk and the Return Using A Hybrid Model

I'm sure you know how to make good deals. So what I want to talk about is what are some of the things that you want to think about the risk? Well, we're small partners, small boutique shops, we acquire mature assets and operate and monetize it on exit, so maybe drill some wells or do some ESP. But the thing I learned quickly-- we never enjoy average wealth-- just the spread. So that's the way to think about risk.

I'm a geophysicist, so how many others are geophysicists here? OK, good. Thank you. But I'd like to quote the geologist of all geologists, Wallace Pratt. He said that oil is finally in the minds of men. In fact, he added on his 91st birthday-- and women. But if he were still here today, maybe given the data that we'd have access to-- all these production information, maybe learning with machine-based, machine-learning alone-- he might suggest that.

Now, let's talk about risk. If I substitute learning with flying-- flying with a machine on a 737 MAX 8 beats machine flying alone. I think that's the way I look at it. OK. So bear with me.

Now, what we're going to go through in the next few slides is just look at risk, but how do people think about it, but particularly if we want to get a better handle on it? Use the granularity of data, but also we want to be transparent. We don't know-- just trust me, and do that, and do this, and put it all together in an actionable wisdom, so we just make better-informed decisions.

Now, there are a lot of books you can read, but this one you read over the weekend-- 138 pages. Bethany was the same author that wrote about Enron. She was a financial analyst at Goldman Sachs for a while, so she knew what she was talking about. But if I can help you buy just the book in one word-- it's about volatility. It's not the price of oil. I mean, it goes up and down. I mean, December was $44. And a couple of days ago, $60-some. But that's not what she's talking about. She's talking about some of the things inherent in our industry, or whether you rush in to go, and it's quite expensive-- acreage and drill and maybe you can really make it work.

But let's see some of the quotes from there. You can read it. But one of the things she pointed out was the valuation was based on multiple-- not multiple profits, as we all now learn, but based on the multiple of acreage, which didn't work quite well.

All right. So this is one that perhaps gets some attention-- based on Goldman Sachs-- that in order to keep the same production level from the conventional, you'd put a lot more capital to work. So let's say 2018, maybe about $36 million, $38 million-- 2023 project about a little over $56 billion. But if you use the same cumulative average growth rates-- could be a huge amount by 2040.

So I don't know, if by 2050, we can switch to renewable, but it's a tremendous amount to keep it at the current level-- not really about growing. So what is sustainable? Not as you're a better judge than that.

Now, what I want to talk about is something about how do we use data. We use a lot of data, but how? Well, as far as this other book talks about, Bridgewater, I guess they run very successful hedge funds, but they really built a database, and they use algorithms-- go through data. So they come up with-- a way to look at it is always continuous learning, and so forth-- how to make better decisions for the next round.

So the algorithm-- but the data-- now, everybody talks about Google, Amazon, and so forth, but I think particularly for oil and gas, we do have some data, but we need to think about it differently. The data, for example, you heard about type curves, and here's one. Just, typically, we get all kinds of curves, but we might average them out. We kind of have one representative of an area. But that's done many, many times. So later you'll find out we've got better ways to use machine learning to help you do that.

But having said that-- so that's a starting point. You get a type curve. And then the data granularity-- so let's talk about spread-- this is the intrinsic volatility in the initial production. Now, whatever the reason-- it so happened this is in Eagle Ford-- you can look at the spread, you say, hey, well, I just kind of fit three different distributions. These days, there's 80-some of these you can fit. But if you look at the purple dash line, that's the normal distribution. That's where we say, hey, we took the average, we have some feel for if we drill enough wells, we'll hit that average. But if you drill only a few wells, it might not work. But the actual distribution looks more like a lot normal, so kind of skewed.

Now, geologists here-- what does it remind you-- the lot normal distribution? Well, the grain size, the free reserves-- a lot of not normal in-- mother nature doesn't play dice. This is not a normal distribution. So what it really means-- the first thing is that if you take an average of something-- you say, hey, that's a P50. No, that's not a P50. In fact, let's say these are three areas you look at in Eagle Ford-- a lot of different levels of production, productivity, but in any case, they all showed some kind of skewed distribution.

So the first thing is that, hey, when we look at risk or whatever the kind of when we want to get a handle on the return versus risk, maybe we should think about how to incorporate the spread in there. OK. So just an example, so let's say I go in, I say, well, I don't know, but I can put in-- I have $100 million. I put 1/3, 1/3, 1/3, to each equal split into those assets. And then I can-- based on the decline curve-- I can do well because I get a distribution of what that spread of the return might be. All right?

Now, looking at this spread is one thing, but how do I make sense of it? So I borrowed a concept from the value-at-risk from the financial industry. That's what people use to determine this with a bank or institution is safe under certain scenarios.

But one other simple interpretation is this-- in the middle, the black dash line-- that's, let's say 50% of time-- I might-- for this portfolio-- I might lose no more than 10% of my investment. And you go all the way to the left, you say, well, about 99% of the time, I would not lose any more than 82%. So it gives you a feel for, hey, if I want to do some hedging, I say, well, I don't want to lose more than 10% so maybe hedge 50% of my investment or 50% of my production volume.

But if I want to sleep like a baby maybe-- I want to hedge more because I push it to the far left, but of course, it's like insurance-- I want more coverage, I have to pay more. But, that's again, one way to look at if you look in an area-- look at the distribution of production, you can have a way to gauge what the spread is. That's just one scenario.

What if there's some other combination? So, for example, maybe for these three assets, I might have different amounts that I put in. So if you go through the scenario assimilation-- so each one of these spreads, then I say, hey, I have the so-called variance and the expected return, and I plot it on this with-- in this case, do 2,500 assimilations, so I get this little diagram. But what's important is that in terms of the expected return versus expected volatility, which is the risk, the spread, it's not random. It has this little shape. It's like, you think of, it's like a little bow-shape if you turn it sideways. But the key thing is that the end-flow-- the edge of that is what some of the financial world calls the "efficient frontier," as some of you probably have used them in your company.

But what it really means is that for the allocation along that efficient frontier for a given return that you seek, there's a portfolio allocation of those assets that you can achieve with minimum volatility. So where the black arrow points to, that particular combination-- maybe half of that-- goes into Asset A, 8% goes into B, 41% goes into C, and then you get about-- so the return's roughly about 12%-- I mean, about less than-- slightly less than 10%, but your volatility also is less than if you go to the right.

But if you really want to go with minimum volatility sometime, you might say, hey, I sacrificed on the return. Now, this is based on average well cost and so forth. So if you have the ability to improve the operating efficiency, lower your cost, then you basically can move the thing up towards the upper left quadrant. That's what you want.

OK. So I basically split this into four quadrants-- the upper-left, upper-right, lower-left, lower-right, so in all that, you get a sense of what that really means. I borrowed this concept from Boston Consulting Group, the Matrix Unit. Lower-left-- you don't want to go there. Those are docs that, hey, they're low return, low volatility, no chance of really turning around.

The upper-right is the stars. It's good return-- good risk-adjusted return. But what you see is that they're all going to-- they are mostly the mirror image-- the inverse of each other when you go diagonal. So, again, it gives you a sense of, hey, how do I think about allocating my capital to work? Now, there's only for 3. Now, you can do it for 80, you can do it for 100 and so forth, but it's the same thing.

So that's using the granularity of data-- using all the distribution of the initial production and giving you a sense of-- but still using the type curve. So we'll come back to that. Now, as far as the data, we want transparent process. So if somebody comes in, hey, this is a really good opportunity, here's a return. But if a single number-- how to think about it. So here we're going to talk about it just as an example.

If you look at the stock market or oil prices, today-- what they were-- and over time, you have one variable. When you have a time series, you can fit the trend. You can do some prediction or forecast. But if you have different-- more than one variable, then you still can do that.

So in the machine-learning world, it's a multivariate regression. But in the simple case where we want to think about in this application, hey, there's some opportunity at Eagle Ford. I can go and restimulate, refract the well, and maybe I'll get some decent uplift. So here we say, hey, look for those either binary, yes, or no-- some places that I might get a before and after four times, basically, 300% uplift and where but those places are.

Now, the features that you go in here is the pop-ins, the volume, the fluids, and the chemicals, and the thing that you want to predict or you want to use as criteria is the delta of initial production before and after. So the vendor-- since we don't have the magic sauce-- so the vendor is a proxy for that-- so there are about 12 features here and about 3,850 wells. So in green are the ones, that, hey, we get good so-called test results that predict or forecast about-- the red is off. But this is not about how good this is, but it's just a way to say, hey, there's a way to use the algorithm with the data coming from frac-focused data and be able to look at the opportunity and start thinking about is that an area that's worth taking a second look.

A lot of geoscientists, when we talk of models, we think of earth models-- the properties associated with this at every location, the velocity or the brittleness, the permeability, porosity-- but in the machine learning, when you talk about really models-- data models or some kind of weight associated with particular features whether that particular one is the pop-ins or the volume of fluid and so forth. So we'll keep that in mind. So what we're really interested is, we think about the decline curve for so much of the data, but are there some ways that I can do better and put all this into work?

All right. So, again, instead of just algorithms, there's some first principle. So here's, I think this is another one of those-- my go-to book-- I think it was published about 1990, and it's out of print, but if you can get it from a used bookstore, it's still worth taking a look. In there, there's a lot of talk about risk but how we look at the opportunity. So, today, we just simply take some of those ideas and put it to an algorithm and works. So here's an example. It's a type curve, and in the current-- there are different maybe modified the Ops equation that lets you look at the different fittings. And so you can do that and play with it, and so you can get a fit. But the key is, hey, maybe I can think about what can machine learning do with this information?

So here's one example. And in this example, we take four of the observations and use it to predict a fifth. So think about it real long. I get all his information. I organize it. So I can train it, let it come back and see how well it does.

So you can see in the orange-- this curve is the modified up equation. So it's like if I look at it, I try to fit it, and, hey, I get a good fit, so I got a pretty good model. But if I go to any one of the wells, you start getting data in, and I will have to do this again, right? So that might be a little bit-- that's OK for 1/2 a dozen wells, but if you've got thousands of wells, that might be a little bit difficult. So on the other hand, in the Neural Network application we can take the information-- maybe the first six months or so forth of the production-- and then we run through this, and we'll be able to train this and be able to predict.

So what we want to show is that the Neural Network-- the yellow band is the input data where we will use for training, and then the rest is how it would use to predict and forecast. So you can see the green-dashed line does fit the actual solid type curve reasonably well. The orange is an example-- just a moving average. So you can see-- just get a flavor of what it's doing. But it's better than moving averages because it keeps track of the detail.

Just for information, the Neural Network is the straightforward, simple implementation and a more powerful one you might come across in literature. This recurrent Neural Network is more using long and short-term memory. So it's just more detailed. It keeps track of the same input, but it keeps track of the case. So either way, there's two to automate that whole process, but it's transparent. I can run through it-- could be thousands of wells, and I can get things done.

So let's go back to the principle. I'd like to point out for these unconventional wells because the client is significant. So what I'd like to-- sometimes look at the half-life-- that means half of the production is whatever's captured within the first 8 or 10 or 11 months. But what it means is that if we're doing a restimulating with fracking to move the needle on well economics, then we have to do it before the half-life.

So let's go back to why we want to start having a tool like the Neural Network so I can capture from early production and start projecting and look at my remaining reserves in place-- recoverable the ow, ow, ow, so we can see, hey, there's some opportunity. We can do it before it gets past this economic half-life.

All right. So these portfolio analytics-- at least it's in addition to perform the allocation, and how do we do it better? Just a kind of quick switch. I just want to show the same two where here is the hot energy. And on the website, you can show they have the four major stocks. And if you look at that, you say, hey, that's the efficient frontier that the yellow star is the minimum portfolio, minimum variance, and the star is the best risk-adjusted return.

So that's the four stocks. But if I add one of the independents, it might change and spread it out because there's a lot more volatility of stock prices. But the underlying thing is the same, that you can get a feel for how you allocate. Now, that's from a pure, simple stock portfolio. But if I take the same approach, look at the production data-- for example, on the left, I have the production data from the four basins-- they worked on 2014, '15-- back in Eagle Ford, Marcellus and Permian-- you get a spread. You get the efficient frontier envelope. On the right-- but if I drop back and I look at the Permian-- the Eagle Ford, Marcellus, I got-- well, the maximum risk-adjust return the red star and the yellow star-- they're still about the same in place, but the spread is smaller. So if you're looking at your portfolio strategies, this might be something to consider as a quick way of maybe deciding how you fine-tune your allocation.

One last thing. Nobody is going to fund these all on their own from banks, or so we always need to think about maybe a mix of equity and debt. So if you have done the portfolio work, the idea is that given whatever the fed rate on the intercept, you can say, hey, wherever that line I draw-- if it's 2%, tangent to the efficient frontier is where the most robust. That means if I have that allocation, I might be able to weather the storm a little better. But if the federal fund rates keep rising, you can think of that little box to go from the green to the red-- it starts getting smaller and smaller.

So at some point, this doesn't make sense to finance these projects. So that's just a quantitative way of, again, using all the data-- the production data-- going through this whole process with the machine-learning perhaps on the decline curve part instead of just the type curve and come up with more robust views of these risk and return profiles. Now, again, if I start doing this, I can even classify it into three bands that have really saved-- the one at risk, the one that's really in trouble, so you can just continuous update. Every day you can look at it. But where I'm going with this is that when it comes to asset pricing, one of the things is I can put together using the quantitative technique with the fundamentals, with a production trend, I can actually come up with a much better handle on where these asset prices may be.

Now, it's all good. But at the end, it's learning with the machine piece-- the machine learning alone. So I want to share with you. And I want to put this out-- get more people with you think about get past using the averages-- that's not your P50-- but one of the things is, if you go to Amazon Skills, you can look up, Well Proved Undeveloped. There's a little thing you can download-- you can play with. And you can give us some average-- some spread. And if you want the P10, P50, P90, assuming it's a lot normal distribution. And you can really gig it out. You can be a P66-- it will give you that too.

So at the end, what I really want to do is to think of developing a digital assistant. What kind of digital assistant do I want? Is it a P10 guy, P50 or P90? I really want him to help me to become the giant of whatever basin I'm playing. This, by the way, that's James Dean, The Giant, in 1956. Thank you. And thank you for your attention.