Discrete AlphaHow it works

Every trade is a buyer and a seller agreeing on a price. Buy orders queue on one side, sell orders on the other; most of the time the two roughly balance and the price moves within a narrow range.

When bad news arrives, that changes. Most participants crowd the same way at the same moment. Sell orders pour in and eat through the queued bids level by level, the buyers who would normally step in pull back, and the price is pushed a long way, often too far.

This is when people panic: they either act too early and buy halfway down, or freeze when the real opportunity finally arrives.

Most directional quantitative strategies, whether trend-following, mean-reversion or factor models, work with time series of price and volume: they start from prices that have already printed and infer the prices to come.

Yet price is a by-product of an auction. Behind a single candlestick lie thousands of bids, offers and trades; on the chart, only four prices and a volume remain: open, high, low and close. In information terms, price is a lossy compression of the auction. It records where trades happened, not how the price was formed.

Extracting features from compressed data, however it is transformed, can only recombine the information that survived. What was lost in compression cannot be recovered from price, however sophisticated the model.

Discrete Alpha studies the auction itself, before the compression. This data is high-dimensional, noisy and irregular, and has mostly been the domain of high-frequency trading, which looks at moves over seconds to minutes. Discrete Alpha denoises and refines it layer by layer to extract information with predictive power for price direction; Chapter 2 draws the structure, and Chapter 3 tests with statistics whether that predictive power is real.

This is also why Discrete Alpha uses no price-based indicators such as moving averages or MACD: they compute on prices after the compression, and by the time they signal, much of the move is gone.

Discrete means acting only at the moments when an edge appears; alpha means every one of those trades has a positive expectation. That is what the name stands for.

Most of the time Discrete Alpha does nothing and waits flat. Of the 8,760 hours in a year it acts in very few, which is why the NAV is flat most of the time and steps up only when an edge appears.

A card counter at blackjack tracks which cards are left in the shoe and raises the bet only when the remaining cards favour them. Discrete Alpha does the same thing: it tracks the orders queued in the market and bets only at the moments that favour us.

We extract, from a chaotic market, the features closest to how prices are actually formed, and turn them into an edge we can bet on again and again.

This is Discrete Alpha’s greatest edge. When a headline lands, we see it when everyone does; but before the news drives the price, the reversal has often already happened in the market. Discrete Alpha finds the turn in liquidity ahead of it.

We do not forecast news, and we do not know which headline is coming. The reversal happens in the market first; the news often arrives later as the explanation.

The case studies show it clearly. The weekend Silicon Valley Bank failed, with the Treasury Secretary ruling out a bailout, Discrete Alpha added heavily 2 hours 15 minutes before the joint backstop. In the extreme fear after the Bybit hack, its buy came 61 hours before the strategic-reserve news.

Hundreds of raw data streams enter the system at once and are computed layer by layer. Each layer performs a single step; further along there is less data and higher information density, so the structure narrows layer by layer like a funnel. Layer by layer this yields decision data, which produces signals across ten input dimensions.

What enters this structure is the raw data of the auction: high-dimensional, noisy and interwoven. The system keeps it as complete as possible while denoising and refining it layer by layer, so the information becomes more concentrated at every step.

The ten dimensions feed five betting logics by regime, then pass through two constraints, long and short. In the end only one decision is left: how much to bet.

Ten dimensions plus the position make an 11-dimensional space. The market is first classified into regimes, each with its own parameter set; within a regime, the position is set by signal strength, adaptive dimensions and several layers of constraints.

There are three regimes: two bet both ways and one holds in a single direction, and each uses different dimensions. Within each layer, height is position.

Set sizing aside: take every signal as it comes, never scale in or out, stake the same amount each time and optimise nothing. Betting with eyes closed like this, 474 bets earned 1.61% each on average, net of fees and funding and without leverage. 68% of them made money, averaging +3.4% when right and −2.1% when wrong, a payoff ratio of 1.6.

If Discrete Alpha had no edge, every bet would have an expectation of zero, and the average of 474 bets would only drift randomly around zero. That drift is measured in standard errors: 2 standard errors is usually enough to call a result statistically significant, and physicists require 5 before announcing a new particle.

Discrete Alpha’s average, +1.61%, sits 8.4 standard errors out. The chance of getting there by luck alone is roughly one in three quadrillion, smaller than flipping a coin 51 times and getting heads every time.

Casinos win over time because every game carries a small edge in their favour: roulette and baccarat have negative expectations for the player. A card counter at blackjack tracks the cards and turns an edge of about 1% their own way.

For every dollar Discrete Alpha stakes it earns 1.6 cents on average, more than the card counter.

Any single bet can win or lose: bet once and the chance of losing money is 32%. But each bet has a positive expectation, and the more bets you make the closer the average comes to it. This is the law of large numbers.

Sampling from the historical returns: over 10 bets the chance of losing money is 9.9%; over 50 bets, 0.2%; over a year of about 114 bets, close to zero.

Was it just a good market? Replace the real entry times of the four signal types with random times in the same year and the same market conditions, keep the exit rules identical, and repeat 10,000 times.

The real entries beat 94% to 99.8% of the random results. Same market, same exits; the only difference is when the bet was placed. The returns come from the signal itself.

Delay every signal by 1 hour up to 7 days before acting on it. The longer the delay, the lower the expectation per bet, which shows the edge comes from fresh information in the signal.

Even at a 7-day delay the expectation is still positive. The edge is strong, not a coincidence.

Was it tuned to the past? Draw 2,001 neighbouring parameter sets at random and run each one independently over the full history.

All 2,001 made money and none blew up. Annualised Sharpe ratios ranged from a minimum of 1.03 to a median of 1.45, all above the 0.92 of holding Bitcoin over the same period. The edge comes from the structure of the signal, not from one carefully tuned set of parameters.

Discrete Alpha made money in the years Bitcoin rose and in the years it fell: Bitcoin ended 2025 lower, and Discrete Alpha was still positive.

Could it stop working? We trade the mechanism that forms prices, not patterns in past charts. As long as markets price by auction, and market makers and large capital keep trading, real supply and demand stays in the market and moves the price.

We trade only Bitcoin: it has the deepest liquidity, and this kind of edge is most stable where liquidity is deepest.

Knowing where the edge is, you still have to decide how much to bet. The Kelly formula answers exactly that: given the odds of winning and the payoff, what fraction of capital to stake each time.

The fraction f = p − (1 − p) ÷ b, where p is the probability of winning and b the payoff (how many times the stake you win; you lose the stake if you lose). The more certain the edge, the larger the best bet; if the result is zero or negative there is no edge, and you do not bet.

Bet too little and capital grows slowly; bet too much and you lose over time, even with an edge every time.

The Kelly curve shows that the best bet depends on the odds of winning and the payoff. But the market keeps changing: the same signal carries different odds in different markets, so the optimum keeps moving too.

Discrete Alpha first identifies which regime the market is in, then finds the best position for that moment within the regime’s 11-dimensional space, from signal strength, adaptive dimensions and several layers of constraints. It takes size when conviction is high, and does nothing when it is not.

Execution is built around passive fills: positions are opened and closed without moving the market or revealing intent.

Execution is part of the edge. With the same 2x leverage from January 2023, filling everything at market would have left the NAV 14% lower by October 2026.

$10,000 invested in January 2023 at 2x leverage was worth $8,402,756 on 13 October 2026; the same $10,000 held in Bitcoin was worth $50,297. The NAV is flat most of the time and steps up only when an edge appears.

Our team comes from UC Berkeley, Imperial College London, UC Irvine and Stanford, mostly in computer science, with physics, mathematics, statistics and artificial intelligence.

Your capital stays in an exchange account in your own name. We have trading permission only, no withdrawal rights. There is no management fee; we share in profits only after the account makes a new high, above the high-water mark.