Have you wondered how much money you can make from crypto trading if you're psychic?

I developed an algorithm called M-FARM that finds the max profit when you can predict prices with 100% accuracy. A thread:

First off, how much money can you make from crypto trading if you're psychic?

The answer: M-FARM achieved ~5.5% return in a 3-hr sample period using historical data. That would translates into a 50% return EVERY DAY.

Why is this useful?

1. We can learn trading from the algorithm by examining the trading behavior of M-FARM on historical data. Combining this with technical and fundamental analysis can reveal patterns for good trading signals and even strategies.

2. Use as a reference metric for backtesting/evaluating trading performance. The M-FARM return can be interpreted as "the best possible return" for a long-only crypto portfolio. (whereas market portfolio return/risk-free return are baselines)

A semi-technical summary:

M-FARM stands for minutely fee-adjusted return maximization. Since I traded readability for a cool acronym, let me explain the detail with 50% technical details + 50% intuition/examples.

Assumptions/restrictions: 

- Long-only: only buy/sell, no shorting/borrowing (otherwise meaningless bc return is infinite)

- Trading on CEX: instant transactions and percentage fees. I'm using FTX US data: only ~17 tokens

- Price taker: our trades do not affect market prices

We are given the following: initial portfolio value (V_0) e.g. $10k; transaction fee rate (τ) e.g. 0.2%; and the price (P) of every asset (a) at every minute (t).

Index trades from 1 to n, then the final portfolio value (V^n) can be condensed into the following expression:

The intuition: With each trade we have to pay some fees, but profit from price increases.

The trade-off: if we trade very frequently in order to capture every upward price tick, the fees will eat a lot of the profit. 

Hence an optimization problem.

Although we have a closed-form expression, this optimization problem is difficult to solve with conventional methods b/c:

1. # of trades (n) is undetermined.

2. The number of variables is too large for most non-linear optimizers. (also technically an integer programming prob)

M-FARM solves this optimization problem by identifying the best trade in terms of minutely return adjusted by fees. The exact definition is as follows: 

Notice the exponent term which is adopted from the classic yield-to-maturity formula used for bonds etc.

Intuitively, 10% return in 30 minutes is better than 15% in one hour. (time is money!)

If you specify a time period (T), M-FARM first finds one (asset, buy_time, sell_time) combination within T with highest M-FAR.

Then it recursively finds the highest M-FAR for the rest of the time period, until every minute within T is occupied with a good trade.

A brief illustration with a 3-hr period:

-> start with search period T=0~180min: best trade is (XXX, 80min, 100min)

-> now search on T=0~80min and T=100~180min;

-> best trade in 0~80min is (YYY, 10min, 15min); best trade in 100-180min is (ZZZ, 115min, 170min)

-> keep going...

Thanks to @ProfPieters for an interesting homework prompt that led to this, and giving insights on how to potentially improve the computational efficiency.

Next steps:

All my code is currently in a messy python notebook screaming for some clean-up. The recursion part currently runs inefficiently at O(n^3) per branch.

I might write M-FARM as a program w/ UI (anyone knows how?), or open source it on a github repo, but at least after the code clean-up

Like/follow/RT if you like the content! I'll be posting more web3 research stuff since I'm a graduate student looking for web3 internship/fulltime positions and could benefit from some twitter exposure.

M-FARM: trading algorithm for psychic traders

The biggest problem in DeFi lending right now is adverse selection imo. 

I'll illustrate it with the classic lemon car example used in my GT class, and share some thoughts about potential solutions. A thread:

1/n
Imagine you're buying a used car but there exist two types of cars at the dealer: peach (good) and lemon (bad, e.g. had an accident). 

Both are polished on the outside so you as the buyer can't tell the difference, but the dealer knows.

2/n
Suppose a peach is worth 20k and a lemon is worth 10k. 

Since there is asymmetric info between you and the dealer, you would price every car at its expected value, which is a price between 20k and 10k and depending on what you think the percentage of each type there is.

3/n
e.g. if you think it's a 50-50 chance then you'd be willing to pay 15k for a car of unknown quality. If the other buyers think in the same way, then 15k becomes the market buying price for used cars.

4/n
The problem: at a market price of 15k, the dealer loses money if they sell a peach. Therefore they'll sell more lemons to make profits. 

Over time, there will be less and less peaches and lemons overflow the market. This is adverse selection induced by asymmetric info.

5/n
Switching back to DeFi lending: there are peach (low default risk) and lemon (high default risk) borrowers in the space. 

Since lenders typically dont know the credit history of borrowers, they will treat everyone as if they have an equal default risk between high and low.

6/n
Lenders will set interest rates, collat ratio, optimize portfolio, etc. according to this framework that everyone has a medium risk.

Then adverse selection happens: over time, good creditors will think the current borrowing cost is too high and quit the protocol.

7/n
Good creditors can switch to another borrowing source where they can prove themselves as a good creditor by KYC to get lower rates.

Bad creditors will remain in the Lemon Protocol (cool name), and the quality of funds deteriorate over time.


8/n
Solution:
On the conceptual level, the most effective solution to asymmetric info is signaling, which is a fancy economics term for 'revealing info'.

9/n
In the used car market most dealers offer Carfax reports. In Tradfi lending, individuals/companies has a credit score from an indep credit rating agency like FICO or Moody's.

As you can see here, signaling is often done thru third party verification processes.

10/n
You might think that the concept of signaling or revealing borrower info sounds somewhat contrary to the promises of DeFi lending, which I agree. 

But as the DeFi ecosystem matures, on-chain data can serve as a good info source too.

11/n
Earlier today I read about an interesting project @arcxmoney. They attempts to establish an on-chain credit score system that determines each user's LTV or collateralization ratio from their on-chain credit history. It's for their users only, but interesting approach.

12/n
Another direction is to use a KYC-like process for credit but with ZKproofs: like showing your income is above a certain range without revealing your bank account. 

This is an ambitious approach, but once solved, it can open doors to undercollaterllization for the masses.

13/n
Of course there's still the long-lasting problem that there is little consequence to just run away from the loan and open a new wallet. The solutions I discussed above can prevent adverse selection, but we still need a good solution on the runaway type of default.

Adverse Selection in DeFi Lending

1/n
Research showcase: I took a research approach to analyze various factors that can potentially predict $MKR returns on different time scales. Those factors include momentum, macro indicators, Internet search trends, other crypto, and network fundamentals.

2/n
Momentum: 
MKR has a strong positive short-term momentum, but the trend often reverses in 5-6 weeks. (Using autoregression with 1-8 weeks lags on weekly data since 2018)

3/n
Google search trends: 
Google searches on MakerDAO positively correlates with the next-week return of MKR. Interestingly, searches on Bitcoin and Ethereum also shows some correlation with t+1 and t+2 returns. (Again using autoregression on weekly data)

4/n
Macroeconomic factors (consumption, production, income): 
Generally no significant correlation, except for growth in durable consumption. Results are robust after controlling for risk-free interest rates. (Log linear regression using monthly data)

5/n
Network fundamentals (address count, transaction count/volume, average transaction value):
Consistent with theoretical prediction, there is significant contemporaneous correlation between MKR return and addresses count, transaction volume, and average transaction value.

6/n
However, that might be reverse causal: people create new addresses and buy/sell more when returns are high. So I looked at this-month fundamentals vs next-month return. Now Addresses count remains a good predictor.

7/n
Technical ratios (NVT, MVRV):
I found no significant correlation b/w MKR returns and its NVT or MVRV. This could be because the delegation mechanism making transaction data fuzzy, or simply that transaction-based ratios doesn't mean much to MKR as a gov token.

8/n
The tables for the last two (fundamentals and ratios) are too big to attach in the thread. They are in the full report here: https://drive.google.com/file/d/1-CPcANp4NJxxyAsbVCLfaAvYLoGqkqLk/view

9/n
Trading MKR:
An effective trading strategy on MKR should pay close attention to its user base, voting-related activities, and search popularity in the past 1-2 weeks. Be cautious to hold any position for over 4 weeks since there's a potential reverse momentum.

10/n
It can be speculated that DAI's TVL, and its major collaterals such as USDC and BTC, can also have an impact on the demand for MKR. Since MKR resides within the Ethereum ecosystem, ETH returns and related events (e.g. the merge) will also impact MKR’s performance.

11/n
This report is my final project for @ProfPieters's finmath crypto class @UChicago. Special thanks to her for teaching us lots of useful research methodologies in crypto, including the one I'm using here, originally outlined by Liu and Tsyvinski (2021).
@MakerDAO #MakerDAO

M-FARM: trading algorithm for psychic traders

Adverse Selection in DeFi Lending

MKR Research Project Showcase