|
Editor's Note: Today is the grand finale of the Hedgeye Investing Summit. Tune in to hear from investing veterans Daniel Lacalle, Tracy Shuchart and Art Hyde. Click HERE for access. |
“I think the danger of AI is much greater than the danger of nuclear warheads by a lot, and nobody would suggest we allow the world to just build nuclear warheads if they want, that would be insane. And mark my words: AI is far more dangerous than nukes.” - Elon Musk
We are releasing a brand-new product today called HedgAI Signals. These signals are generated by our proprietary “HedgAI” Artificial Intelligence (A.I.). We will be providing these signals for free for the next few months (estimated early 2024), until they become a subscription product.
Use THIS LINK to opt into receiving the signals daily.
You do not have to be a subscriber. Feel free to pass along this link to anyone you like. Tomorrow at 12:00pm ET, Daryl Jones and I will have a live webcast discussing these signals in greater detail, which you will receive access to when you sign up. I will start off this Early Look by describing the technology behind these signals then go into the output you will receive daily.
If no one else has done so, allow me to welcome you to the phase of human history that can (and will) alter the course of every generation to come. What’s so special about this phase? It is something called Artificial Intelligence aka A.I.
Pretty big claim. Let’s get into it.
One subset of A.I. is artificial neural networks. They are a type of learning inspired by the structure of a human brain.
Have you ever thought about how you learned to do simple tasks? For example, let’s look at how we learned how to catch objects. Think about an object being thrown at you when you were young. Over time, your brain came up with two different solutions to this problem, catch the object or dodge the object. I can guarantee the first time a teddy bear was thrown at you it just hit you. No harm no foul, but your brain started to come up with a solution and started the first process of learning (backpropagation), which is not liking an outcome (getting hit with an object) and changing future behaviors to try and solve for a better outcome.
From that moment on, your brain continually trained itself to take in the inputs that best allowed you to catch objects. These inputs flow through a series of neurons, and tells your body what to do next. In other words, if I threw a pencil at you, your brain would identify:
- How large the object is
- How fast the object is coming at you, and
- How far away the object is from you.
It will then take those inputs and decide whether to catch/not catch the object. I learned this as a kid when someone threw a medicine ball at me as a kid. I tried to catch it. It was too heavy. It went through my arms and landed on my foot. I couldn’t walk on that foot for a couple days. My brain adjusted itself and learned not to attempt catching medicine balls.
This task of catching objects takes all of us years to learn (I’m 26, so for me, 26 years to continually learn this task). What if I could give you all the positive and negative experiences of previous generations in just one second? What if I also told you this idea (of neural networks) could be applied to the stock market?
Welcome to HedgAI Signals, where our goal is to provide valuable signals to subscribers so they can make money.
Neural networks are often discussed as the most efficient way to deal with unstructured data. The downside is it requires massive amounts of data to train (learn). But how much data are we talking about?
Let’s do another fun exercise with our Apple (AAPL) model and how we came up with the final signal you see today.
For our Apple model, there are ~125 inputs that we believe are the best predictors of the future stock price of Apple. We then put these inputs through a series of weights and biases (not your typical definition of bias). To simplify these weights, think of weighted averages (2*x)+(3*y)=1 where the weight of x is 0.26 and the weight of y is 0.16. These weights in neural networks continuously adjust in order to find the most optimal distribution based on specific situations. How many different weights and biases does one sample Apple model have? 4,295,197,697 (4.3B).
Let’s assume the model makes one adjustment to all weights (in reality this is an oversimplification) per each batch- our Apple model has 364 batches. That creates 1,565,062,660,844 (1.6T) different models, with different weights for each input, that can be used to predict Apple. Almost like we created 1.6T alternate realities that were tested every trading day from 11/22/1993 until this moment in 2023. But it doesn’t stop there. Let’s say we did this process 16,000 times. Well now, you just created 25,041,002,573,510,000 (25 quadrillion) models.
Cool. But that seems like a lot of different calculations. How long does something like that take to create? Well, when you employ something called a GPU, which has thousands of cores, as compared to your CPU which probably has ~10 cores and is only utilizing one of them 99% of the time you greatly reduce processing time. It will take a computer ~3 days to complete all of those different models. How long will it take a human to do this and properly optimize each weight? If someone wants to try it out, let me know, but I’m going to assume a lifetime or two.
So what have we done? We just employed a computer to simulate itself as if it started trading Apple (hypothetical trades) in 1993 and then continuously allowed the computer to make updates to itself with the objective of making money.
Why does Elon think this is such a danger? For starters, we just lowered the learning time from a lifetime to 3 days. Who has the power to lower this processing time? Anyone with access to large and expensive computers, aka the wealthy. Is it possible to influence the inputs and outputs to get an answer you (as a human) want? Yes, absolutely. Let’s say I had a newspaper. I can ensure the news article generated by the A.I. only ever makes articles with a political bias to one side. But think larger... what if A.I. controlled the entire internet and you were only able to search using an A.I. Algo?
If the creators of said Algo wanted to erase all mentions of the word “George Washington” well, they can do that. Then, future generations using this search engine will no longer be able to learn who George Washington was, or what he did. So, I will ask, who do you trust to control A.I. and ultimately what our future generations will use to learn from?
Back to the Global Macro Grind…
Welcome to Hedgeye’s version of an unbiased neural network.
Our product, HedgAI Signals, are the signals generated from Hedgeye’s A.I. models. They show you a purely math-based signal, utilizing the Hedgeye #process of what the model believes will happen in the future.
The main output you will see is in the Chart of the Day. Every day, using refreshed market data, each model will provide updated signals for you. The three possible signals are “Long”, “Short”, and “Wait and Watch”.
Each Algo that creates a signal is completely unique to the specific asset. This means the inputs are unique, and the training process is specifically designed for that singular asset. We did NOT create a general A.I. algo then deploy it across multiple assets. These are uniquely designed algos, essentially an artificial brain whose only job is to create signals on a specific asset.
In addition, these algos are completely unbiased (classical definition). We don’t get paid to push a certain narrative. I don’t care if we are long or short any of these signals as long as it is correct. This means the only inputs in these algos are the quantitative data, no narratives, no opinions, just a bunch of 0s and 1s. It also means we have zero idea what tomorrow’s signals are going to be until the algo runs for that day.
Whatever it says, is what it says, plain and simple. It’s just math.
A little more on these signals. Each signal is trained to predict 15 trading days into the future, although most algos hold positions for between 3-10 trading days (lower than the trade duration). This enables the signals to update with new data. Imagine oil goes down 50% tomorrow, your perceptions of the market will also update. Holding your view before said event and not adjusting with new information is just not how investing works.
The average holding period is important. These algos do not consider the trade duration, the trend duration, or the Risk Ranges on a trade and trend basis. They are completely separate. But, average HedgAI holding periods being less than the trade range duration puts another tool in your own quantitative methodology.
Looking at today’s output, you can create an A.I. short term view of the market. For example: Crude Oil, Natural Gas, Copper, and Corn are all signaling “Long” which could equate to a short term view of Long inflation. A Long of Nasdaq and XLF signal could be an oversold signal at the index level with an over bought signal on Meta and Microsoft, the single stock level.
We also plan on releasing the following algo signals over the next three months (for free), a total of 51 daily signals, which we believe will only help our signaling process on global macro. XLY, XLV, XLK, XLP, XLI, XLB, XLE, XLU, XRT, Dollar, UST 30yr Yield, UST 10yr Yield, UST 2yr Yield, HYG, Russell, Shanghai, Nikkei, DAX, Brent Oil, Silver, Sugar, Steel, Uranium, Lithium, Wheat, Palladium, Platinum, Cotton, Coffee, EUR/USD, USD/YEN, GBP/USD, and CAD/USD.
The Chart of the Day above shows something called a “Threshold” and columns for “Today”, “1 Day Ago”, “2 Days Ago”, “3 Days Ago”. All of these numbers are referring to Z Scores. Specifically, the published Z Score for “Today” is a trailing twelve-month (TTM) Z Score of the Algo’s predictions. We then provide the previous day’s published TTM Z Scores in the following columns.
Each asset has a unique threshold that marks the point where signals are opened and closed.
- The Algo opens a Long signal if the TTM Z Score goes above the published threshold.
- The Algo closes a Long signal when the TTM Z Score dips back below the same published threshold.
- The algorithm opens a Short signal if the TTM Z Score goes below the additive inverse of the published threshold (i.e. if a threshold is 1.0, a Short signal is opened when the TTM Z Score hits -1.0).
- The Short signal is closed when the TTM Z Score goes back above the same additive inverse of the published threshold.
- 'Wait and Watch' is an output generated by the Algo when the TTM Z Score is neither above nor below the published threshold or its additive inverse.
- During 'Wait and Watch' periods, the Algo assumes no signal (Long or Short) in the asset.
One question we get all the time is: if the Algo’s TTM Z Score is large, does that mean it has higher conviction? The answer is that not all Algos are the same and the thresholds being used are the ones we tested for each algo and believe provide the best results for that Algo. However, since each Algo is unique from the others, multiple Algos showing similar outputs in a similar asset class indicates a stronger signal.
This is a great time for me to remind everyone to please read our disclaimer on this product. We have back tested each Algo against years of historical market data and have been pleased with the results. We look forward to applying #timestamps to future iterations. These are A.I. generated signals and not a certain outcome. Neither Hedgeye, nor any market prognosticator, can know for certain the actual price range an investable instrument can or will trade at in the future.
Another question we get: does this use the Risk Ranges and how does it relate to the Risk Ranges? No, these algos do not utilize the Risk Ranges. These algos are completely separate from the Risk Ranges. While the algos were tested to make sure they work on their own, they can be used to augment already existing processes.
Using the chart below of AMZN as an example (where no real trades were placed by the Algo) we overlayed the trade Risk Ranges with the hypothetical signals of AMZN.
These algos were designed to capture outsized moves where a human perceives a "falling knife." An algo’s ability to see just 0s and 1s becomes an advantage due to seeing this data much differently than a human eye.
As it relates to the AMZN chart above, the red arrow in April demonstrates this. Where using the Risk Ranges one may have covered the position, the algo would have signaled a short. In addition, looking at the Green arrow in April, it looked like AMZN was never going to bottom. However, the algo signaled long close to the bottom of the move. But like all processes, nothing is 100% in trading (unless you are Bernie Madoff). You can see in June the black arrow where the algo had no signals (Wait and Watch).
Don’t worry, you don’t have to remember all of this. The bottom of each iteration will explain all of this again, how to use this product (and the disclaimer). I highly recommend reading both to answer any other questions you may have. Also keep an eye out for a video recording explaining exactly what you will see.
Again, use THIS LINK to receive this publication daily (before market open). When you sign up, you'll also receive access to a free webcast tomorrow (10/13) at 12:00pm ET to learn more about this new product. We will be #timestamping results in future publications as well as adding more Algos to the publication for free.
Immediate-term Risk Range™ Signal with @Hedgeye TREND signal in brackets
UST 30yr Yield 4.63-5.02% (bullish)
UST 10yr Yield 4.50-4.85% (bullish)
UST 2yr Yield 4.95-5.19% (bullish)
High Yield (HYG) 72.03-73.65 (bearish)
SPX 4 (bearish)
NASDAQ 12,978-13,729 (bearish)
RUT 1 (bearish)
Tech (XLK) 161-172 (bearish)
Energy (XLE) 84.30-90.99 (bullish)
Utilities (XLU) 55.16-60.32 (bearish)
VIX 15.73-20.78 (bullish)
USD 105.11-107.01 (bullish)
EUR/USD 1.043-1.065 (bearish)
Oil (WTI) 82.11-91.97 (bullish)
Nat Gas 2.97-3.58 (bullish)
Gold 1 (bullish)
Copper 3.51-3.73 (bearish)
AAPL 169-182 (bearish)
