About

Justin Chuk
Business Administration, Spieker Program
Minor in Data Science
UC Berkeley, Haas School of Business
Class of 2029

Hong Kong → England → Vancouver → Berkeley

I am from Vancouver, British Columbia. I grew up in Hong Kong, went to school in England, and finished high school in Vancouver. I am at Berkeley now, studying business and data science. My work so far has run across energy research, private equity, and quantitative research. Different rooms, one question: how a price gets set, and what the person on the other side of it is being paid to take on. The answer is usually a constraint that somebody cannot trade around, and that is the part I want to work on.

I opened my first account in December 2022. I spent the better part of a year taking positions on stories I had read enough times to mistake for analysis. The uncomfortable part was never losing money. It was that I could not tell the difference between the calls I got right and the ones that went right anyway.

What came after was slower and better. I moved to equities, traded mean reversion by hand, and put the profits into dividend payers and royalty names on a value framework, reinvesting the distributions. It worked, more or less, and it exposed the real problem: every decision still ran through me on the day I made it. I wanted to know whether the process could be separated from the mood. That question is behind everything systematic I have done since, including a study that spent sixteen years of data telling me the answer was mostly no.

At Berkeley I do energy research under Professor Robert Edelstein. Over three phases I wrote on global energy trade flows and chokepoints, on electricity demand growth and the infrastructure constraining it, and on why AI data centres and proof-of-work mining are treated as one demand category when they behave like opposites. Those are summarised on the Energy page.

Alongside school I have worked in private equity at a lower middle market industrial search fund, sourcing and underwriting acquisitions in B2B industrial services, and in quantitative research at a systematic fund in London, building the data infrastructure behind a research pipeline and testing statistical relationships across a universe of listed equities.

The research and the personal book arrived at the same place from different directions. A forecast is the easy half of the work and usually the wrong half. What decides the outcome is the constraint underneath it: a pipeline that cannot reverse, a transmission line that takes eight years, a pool too thin to absorb the order. That is what I look for first, and it is why the energy research and the trading have turned out to be the same job.

Writing

Notes on energy, macro, method, and companies. Levels quoted are as of the date shown and are not current.

Grouped by when the work was done. Publication date shown on each entry.

Eric Hamber Secondary's independent student newspaper, Vancouver. Written in grades 10 and 11.

The edge that only exists after 2019

Updated 28 August 2026 · Method · Study run 2025, re-run on full data 2026

Sixteen years of NQ minute data says the opening range breakout has no edge before 2020 and a real one after. Most of the work was spent trying to break that conclusion.

The data

Databento GLBX.MDP3, one-minute OHLCV, parent symbol NQ.FUT, 2010-06-06 to 2026-07-03. 7,197,763 bars across 71 outright contract months. SHA-256 of the data file inside the delivered archive is 7160b4cf, recorded in my audit log alongside the job ID.

The vendor's condition file flags 13 missing and 28 degraded dates out of 5,054. All 41 are excluded. The missing days follow a pattern: every one is a Saturday, seven days apart, unbroken from 14 February to 9 May 2026. That is a systematic gap in vendor coverage, not random loss. Saturdays carry no RTH session so it does not touch this study, but a rule that treated them as zero-volume days would have quietly mislabelled thirteen sessions.

How it was built

The rules are short. Mark the high and low of the 09:30 to 10:00 ET range, take the long when a five-minute bar closes above the range high, fill at the open of the next one-minute bar. Stop at the range low. Flatten at 14:00 ET. A break below the range low does not disarm the setup.

Of 4,114 sessions with front-month RTH data: 696 skipped on the range filters, 1,283 never triggered, 21 dropped for incomplete data. That leaves 2,114 trades. This run includes FOMC decision days. The frozen live spec skips them, so the study describes a slightly wider rule set than the one I actually trade. In the reproduction run, excluding the 136 FOMC sessions in range removes 57 trades, from 2,123 to 2,066.

The result

Net of costs, $20 per index point. Historical max drawdown −$18,522 on the post-2020 half. Average win +$1,180, average loss −$1,186 on the post-2020 half, from the reproduction run.

Year by year

The split-half framing hides how consistent the break is. Every year from 2020 onward is positive, and every one of them beats every pre-2020 year except 2010, which has 65 trades and should not carry weight.

Reproduction run, gross R, NQ one lot. Yearly samples are small and none of these intervals is shown; read the direction, not the individual years.

The first half's confidence interval straddles zero. The second half's does not. Averaging the two halves together produces a number that describes neither one.

The check that nearly killed it

Dollar P&L is not comparable across sixteen years of NQ. Mean entry price was 4,445 in the first half and 17,078 in the second. A fixed point value means the same percentage move pays roughly four times more in 2024 than in 2012. That alone could manufacture the entire regime story.

So I re-ran it in risk-normalised terms, as an R-multiple where R is the distance from entry to the stop.

Original run. The August 2026 reproduction gives −0.0033 for 2010 to 2019 and +0.0858 for 2020 to 2026, so the pre-2020 point estimate is negative rather than positive. That is the direction the year table above reflects. The verdicts are unchanged and the reproduction makes the first half look worse, not better.

It survives. Mean opening range was 0.444% of price in the first half against 0.520% in the second, so volatility widened somewhat, but not nearly enough to explain a result that roughly quadruples in risk-adjusted terms.

Splitting the good half again

The obvious objection to any regime claim is that you found a lucky stretch. So I split the post-2020 sample down the middle.

Same sign, near-identical magnitude, both intervals clear of zero. This is the strongest evidence in the study. A lucky stretch would not usually split into two halves that agree this closely.

Why the break sits where it does

A regime-dependent result is only worth keeping if you can name a mechanism. Four things changed the structure of the Nasdaq open and all four landed in the same window.

I would put most weight on 0DTE. When a large share of open interest expires the same day, dealer gamma hedging becomes flow that must happen intraday rather than across weeks, and it is heaviest where the move is largest. A breakout is a move. That gives the setup the thing most technical patterns lack, which is a counterparty who is not choosing to be there.

All four are plausible and none is tested. I can point at them; I have not shown that any of them causes the result.

Fifteen ways to make it better, none of which did

I tested roughly fifteen families of enhancement: range filters, volume conditions, time-of-day cutoffs, volatility regimes, trend overlays, a break-and-reclaim variant, and a short mirror. Every one failed the pre-registered criteria. Two failed instructively.

The reclaim variant, waiting for a break below the range then a recovery, appeared to work. It only survived on days that were already baseline winners, and it entered them with a worse stop. It was not finding a signal. It was selecting for days that had already confirmed.

The short mirror failed for a related reason, and failed cleanly: −$7.39 a trade, both halves negative. Shorting the days that broke down and later reversed meant systematically shorting the baseline's best long days, at a cost of about $312 per trade on the overlapping sessions. The two sides were not independent signals. They were the same sessions read backwards.

Fifteen tests at a five percent threshold with all nulls true produce 0.75 expected false positives and about a fifty-four percent chance of at least one. Reporting only the variant that passed would mean reporting a coin flip as a discovery, so the count goes in the writeup including the abandoned tests.

The test I should have run first

A long-only rule that works, whose exact mirror loses in both halves, over a period when the index roughly tripled, has an obvious alternative explanation. Maybe the breakout does nothing and I am being paid for being long the Nasdaq in the afternoon on days that happened to go up. Nothing in fifteen enhancement families tested that, because none of them touched the direction or the holding window. They all took the breakout for granted and argued about filters.

So: same filtered days, same 10:00 to 14:00 window, same stop, same target, same sizing, and no breakout condition at all. Just buy at 10:00. This test is not in the fifteen counted above: it is a null comparison the strategy could only fail, not a variant that could have been selected, so it adds no selection opportunity.

Reproduction harness, MNQ with volatility-target sizing, FOMC days excluded. Not the NQ one-lot figures used elsewhere on this page.

Unconditional afternoon exposure is worth nothing on these days. Both intervals span zero and both profit factors round to one. Whatever the breakout rule is doing, it is not collecting drift, and the beta explanation does not survive.

The decomposition is the part I did not expect.

The breakout fires on 64% of filtered days and separates them almost cleanly. That is a real selection effect and it is where the edge lives. But it also means that by the time the condition confirms, more than half the day's move has already happened: entering after the break returns $26.19 against the $60.43 available to someone who had been long since 10:00 on those same sessions.

That gap is not an opportunity and I want to be explicit about why, because it is the most tempting number in this study. Capturing it would require knowing at 10:00 which days will break out later, which is the entire problem. It is a decomposition computed with hindsight, not a strategy. Any rule that tried to enter earlier would need a new signal, tested from scratch, and it would be a parameter change to a frozen spec.

It also has a familiar shape. It is the reclaim variant's failure written from the other side: a condition that looks predictive and is actually confirmatory. I found that twice in the same study, once by accident, and I only found it the second time because I finally compared the strategy to doing nothing.

The decision

I froze the specification. No parameter changes, no new filters, no post-hoc exclusions until the live test below is finished. The rules for what counts as passing were written before the first live trade, so they cannot be moved once I can see how it is going.

Freezing feels like giving up on making it better. It is not. If I keep tuning, I will eventually find a variant that tests well and trades badly, and I will not be able to tell the difference, because by then the data would have been used too many times to say anything.

Testing it for real

A backtest is a claim about what would have happened. The only way to find out whether it survives contact with an actual order book is to send orders, so the frozen strategy is now running live at the smallest size the contract allows, through an evaluation account at a proprietary trading firm. That structure suits the question: the firm puts up the capital against a set of drawdown rules, which caps what a bad answer costs me while still producing real fills, real slippage, and real rejections rather than simulated ones.

What I am measuring is the gap between the model and reality. Did the order fill where the backtest assumed it would, or a tick worse. Did the stop slip further than the two ticks I modelled. Did a trade get missed entirely because of latency or a platform error. Those differences are invisible in historical data and they are the usual reason a study that looked fine stops working when money is attached.

The rules for what counts as passing were written before the first live trade, so they cannot be adjusted once I can see how it is going. If the live results diverge materially from the model, the strategy goes back to diagnosis rather than to another round of parameter tuning, because a result that only holds in backtest is not a result.

What Hormuz bypass capacity actually covers

11 July 2026 · Energy · The strait has been closed since 28 February 2026. The body below is preserved as written; the addendum at the end maps it against the war.

Everyone quotes the bypass number. Almost nobody quotes what it is a percentage of.

The Strait of Hormuz carries roughly 20% of global petroleum liquids and about the same share of LNG trade. Under normal conditions 14–17 mb/d of crude and products move through a channel two miles wide in each direction.

Against that, the entire Gulf has two functioning oil bypass routes.

Six and a half million barrels a day of pipeline against fourteen to seventeen through the water. The gas side is worse: Qatari LNG has essentially no alternative route, with the Dolphin line to the UAE and Oman as the only exception.

The part that gets less attention is where the spare capacity sits. Roughly 3 mb/d of global spare oil production is concentrated in Saudi Arabia and the UAE, which is to say behind the strait. The buffer and the blockage are on the same side of the door.

An earlier version of this piece flagged a ratio I could not reconcile: bypass capacity widely cited as covering about seventeen percent of typical Hormuz flow, when six and a half against fourteen to seventeen is closer to forty. The resolution is definitional. The seventeen percent figure counts unused capacity, capacity not already carrying committed barrels, while the forty percent version counts nameplate. Both are right about different questions. The one that matters in a closure is the unused number, because the committed barrels were already not transiting the strait.

Then it happened

Added 29 August 2026. This piece was published in July written in the hypothetical, and that was a framing mistake, because by July the hypothetical had been the news for four months. The strait closed to commercial traffic on 28 February 2026, and as of this update on 29 August, commercial transit remains largely halted, six months in. If you are reading this later, check the current status; this paragraph is stamped, not live.

Every claim in the table above got tested. The threat-envelope point in the box below proved out first: Petroline's pumping station was hit in April, Fujairah's loading operations were suspended by drone strikes in March, and Yanbu, where Petroline terminates, was attacked as well. The bypass routes sat inside the same war as the strait, which is what the box below said the risk was.

The Qatar row aged the hardest. Iranian strikes on Ras Laffan in March damaged two LNG trains, about 17 percent of export capacity, with repair estimates of three to five years, and QatarEnergy has been under force majeure since March, extended into October as of this update. Zero bypass capacity was not an abstraction. Iraq's southern fields went to a near-total shutdown as storage filled.

One thing the piece understated: capacity is not static under pressure. Saudi Arabia pushed Petroline to around 7 mb/d and kept roughly 60 percent of pre-war exports moving, and the UAE fast-tracked a second Fujairah line. The nameplate table above is now a floor, not a description. What did not change is the shape: Iraq, Kuwait and Qatar still have effectively nothing, and the arithmetic of the piece held where it mattered.

Two digital loads, one label

26 July 2026 · Energy

AI data centres and proof-of-work mining get grouped together constantly. They are close to opposite grid problems.

Both are digital, both are large, both are growing. That is where the similarity ends, and treating the resemblance as the important part produces policy that misses on both.

A grid emergency cannot be resolved by asking a hyperscaler to pause inference. A commercial AI workload is functionally a critical infrastructure load. Mining is the opposite: it will curtail instantly, which sounds useful until you notice that the same price sensitivity means it scales back up at exactly the moment an operator planned for reduced demand.

The measurement is not equally good either. Data centre figures are anchored to disclosed capital expenditure, hardware shipments, and interconnection requests. Bitcoin mining ran an estimated 155–175 TWh in 2024 on a methodology whose published bounds span 80 to 390 TWh. Those two numbers do not deserve the same confidence and should not sit in the same sentence without saying so.

The share is small and the problem is local

Data centres are under ten percent of global electricity demand growth this decade. That sounds manageable until you look at where they sit: 26% of Virginia's electricity, and 23% of Ireland's entire metered supply. In 2024 a minor disturbance in Fairfax County pushed sixty facilities onto backup generation simultaneously, removing about 1,500 MW from the grid in a moment, roughly Boston's entire draw. Concentrated load is a demand problem and a stability problem at once.

The plateau that ended

28 June 2026 · Energy

Advanced-economy electricity demand was flat for thirteen years. Every planning assumption built on that is now wrong.

From roughly 2009 to 2022, electricity demand in advanced economies barely moved. Efficiency gains in lighting, appliances, and industrial processes offset new consumption almost exactly. A generation of utility planners built integrated resource plans around a flat line.

In 2024 global demand grew 4.3%, an increase of 1,080 TWh and double the average annual gain of the prior decade. Emerging economies drove roughly 85% of it, China alone more than half. But the more consequential fact is that advanced economies rose too, led by the United States.

The drivers behind that are not cyclical. Data centres, EV charging, industrial electrification, and cooling all compound year over year. Cooling has a weather component that will fluctuate, but each new air conditioner permanently raises the baseline. The rest is structural growth in systems that have not planned for growth in fifteen years.

Grid length needs to expand roughly 40% by 2035 and annual grid investment needs to roughly double. Planning, permitting and energising new transmission in advanced economies takes five to fifteen years, three to five times a utility-scale solar farm, and transformer lead times have doubled in three years. About 20% of planned data centre projects globally are already at risk of delay for this reason alone.

The near-term consequence is uncomfortable for anyone holding a net-zero target. The IEA expects gas and coal together to meet over 40% of additional data centre demand through 2030, not by preference but because clean capacity cannot be permitted and built at the speed the load is arriving. That projection predates the 2026 Gulf war, which has since taken a fifth of Qatari LNG capacity offline; the gas half of it now runs through a disrupted market.

Tokenization doesn't create liquidity

22 August 2026 · Crypto · Market structure

I started in crypto and my first paper was on how automated market makers price a trade. Both of those make me more sceptical of the tokenization pitch, not less.

I have been following this since 2025. The figures below are current as of August 2026.

The pitch is familiar. Take an asset that trades badly, put a token wrapper on it, and it becomes liquid: fractional, 24/7, instantly settled. Real estate, private credit, pre-IPO equity, anything with a wide spread and a slow close.

The wrapper is real and the settlement improvement is real. The liquidity claim is where it breaks, and the reason is mechanical rather than ideological.

Depth has to come from somewhere

When I compared fee structures and marginal slippage across AMMs, the finding that mattered was not about fees at all. It was that the price you get is a function of pool depth against your trade size, and nothing else in the design changes that. A constant-product pool will quote you a price for any size. It just quotes a worse one as the size rises, and the curve steepens fast once you are large relative to the pool.

Tokenizing an illiquid asset does not add anyone willing to take the other side. It creates a venue. Whether that venue has depth depends on whether a market maker chose to put inventory there, which is a capital allocation decision made by someone weighing spread capture against inventory risk on an asset that was illiquid for a reason.

The numbers, as of now

Tokenized real-world assets sat around $29B on chain in April 2026, up from roughly $7.9B a year earlier. Tokenized Treasuries alone passed $15.3B in May. That growth is genuine and I am not disputing it.

But most of that value does not move. Treasuries and money market funds dominate the total, and those are held for yield, not traded. The category where the liquidity claim actually gets tested is equities, and there the market cap is roughly $2.4B against about $9B of year-to-date on-chain volume.

Two details in that market are worth more than the headline. Roughly 83% of one recent month's tokenized equity volume ran through a single platform, and Solana carries the large majority of on-chain equity transfer volume. An ecosystem where one venue's uptime and compliance posture determines whether the market functions is not a decentralised market. It is a single venue with extra steps.

Fragmentation is the part that compounds

A tokenized share on one chain is generally not fungible with the same exposure on another. Different issuer, different wrapper, different pool. So each new deployment adds an isolated pocket of depth rather than deepening anything that already exists. Aggregate volume rises while the depth available to any individual trade does not.

That shows up in prices. Canton's 2026 tokenization work measured pricing gaps of 1 to 3% for identical assets across chains, and 2 to 5% in friction moving capital between them. In a market with unified depth those gaps get arbitraged to something near zero. That they persist is the measurement of the fragmentation.

The netting problem nobody advertises

This is the part I find most interesting and least discussed. Atomic settlement is sold as pure gain: every trade settles individually and immediately, so counterparty and settlement risk fall. Both true.

But netting is why market making is capital-efficient. A dealer running thousands of offsetting trades through a clearing cycle finances the net exposure, not the gross. Settle every trade independently and that dealer has to prefund or collateralise each one. Gross liquidity demand rises, intraday funding rises, balance sheet usage rises.

So the settlement upgrade makes the economics of providing depth worse, in a market that needed more depth. That is not an argument against atomic settlement. It is an argument that the liquidity benefit and the settlement benefit are pulling against each other, and the marketing counts both.

Where I think it does work

Tokenized Treasuries make sense, and it is worth being precise about why: the value there is collateral mobility and near-instant transfer of a cash-equivalent, not price discovery. Nobody needs a deep secondary market in a T-bill wrapper. They need it to move on a weekend.

The same logic extends to anything held for yield and moved for operational reasons. It does not extend to the assets the pitch leads with. If an asset was illiquid because few people want to own it at the offered price, a token does not change that. It changes where you find out.

The circle has to close somewhere

23 August 2026 · Markets · Figures current as of publication; addendum added 29 August

I spent a semester researching the grid buildout for AI demand. The uncomfortable part was never whether the electricity is available. It is whether anyone is going to pay for the thing the electricity is for.

The structure is now well documented. Nvidia announced an investment of up to $100B in OpenAI to support a data centre buildout equipped with Nvidia chips. OpenAI contracted for roughly $300B of Oracle cloud capacity. Oracle builds those centres with Nvidia hardware. Nvidia books the sale.

Bernstein's Stacy Rasgon flagged the obvious on the day the Nvidia deal landed, writing that it would fuel circular concerns and raise questions about the rationale (Bernstein, Sep 2025). Nvidia participated in more than 50 AI venture deals in 2024 alone, several of which then spent the capital on Nvidia GPUs. The CoreWeave arrangement is the cleanest example: Nvidia invested, supplied the chips, and is among the largest customers.

The unit economics underneath

Circularity only matters if the end demand is not there, so the question is whether the product pays for itself. On the numbers, not yet, and not close.

Read the margin line first. An operating margin of roughly −122% means every incremental dollar of revenue currently costs more than a dollar to serve. Not in aggregate after fixed costs. Per dollar.

The part that gets misread is the efficiency story. Inference cost per query has fallen roughly 95% since GPT-4 launched in early 2023, and gross margin has improved accordingly. That is real progress. It is also being swamped, because volume and compute commitments are growing at least as fast as the per-unit cost is falling. This is the same dynamic I wrote about in the energy research: Google cut median energy per Gemini prompt by a factor of 33 in a year while total consumption rose. Efficiency gains that expand deployment do not reduce the bill.

Where I think the risk actually sits

Not with OpenAI. It has raised enormous sums and can plausibly keep raising them. The exposure is with everyone who has already converted a promise into physical capital.

On 2 February 2026 Oracle issued a statement saying it remained confident in OpenAI's ability to raise funds and meet its commitments. The stock closed down roughly 2.8%. One venture investor described it as bank-run language, which is unkind and not wrong. Reassurance about a counterparty's solvency is information about the counterparty's solvency.

Here is where my own research makes me more worried rather than less. Grid infrastructure is the binding constraint on data centre buildout: transmission takes five to fifteen years in advanced economies and about 20% of planned projects are already at risk of interconnection delay. Utilities, transmission developers, and turbine manufacturers are committing multi-decade capital against load forecasts built from disclosed AI capex.

If the compute demand is durable, that is prudent planning. If a meaningful share of it is a promise financed by the supplier of the equipment, then the stranded asset is not a GPU. GPUs depreciate on a three to five year cycle and someone eventually eats it. The stranded asset is a substation, a transmission line, and a gas turbine ordered against a forecast, with a thirty to forty year life and a regulated ratepayer at the end of it.

Added 29 August 2026. There is a second layer I should have connected when this ran, because it comes from my own Phase 1 work. The gas-turbine bet is now being placed into a wartime gas market. Hormuz has been closed since February, QatarEnergy has been under force majeure since March with roughly 17 percent of its export capacity down on a three-to-five-year repair estimate, and QatarEnergy has been buying US spot cargoes to cover Asian customers. That pull cuts both ways for the turbines in this piece. If the scarcity persists, the fuel under the forecast gets more expensive and the operating case for gas-fired data centre supply gets worse. If the war ends and Qatari trains come back into the middle of the LNG buildout already under construction, the late-decade market loosens sharply, and turbines ordered at scarcity prices run into a glut. Either branch adds variance to a thirty-year asset ordered against a five-year demand forecast, which was the point of the piece before the war made it larger.

What I am not claiming

Not that the technology is useless. I use it daily and the enterprise revenue mix is growing, which is the healthiest signal in the numbers.

Not that this is fraud. Vendor financing in a capital-intensive industry with a scarce input is a defensible way to lock capacity, and calling every instance a bubble is lazy.

What I am claiming is narrower: revenue growing while the operating margin gets worse is not a scaling story, and a structure where the supplier funds the customer makes the demand signal harder to read at exactly the moment other people are pouring concrete against it.

Ed Zitron has been the loudest voice on this and is worth reading, though he is a commentator rather than a source, and I have tried to cite the underlying filings and analyst notes rather than the commentary about them.

Energy

Three research papers written under Professor Robert Edelstein at UC Berkeley, on trade flows and chokepoints, electricity demand growth, and the digital load created by AI and proof-of-work mining, plus the group synthesis report they fed into.

Research

Published research and systematic work.