Qiran (@qrlow) header
Qiran (@qrlow) profile picture

ex-Goldman commodities sales and trading, ex-hedge fund commodities analyst, exploring the intersection of commodities and AI. More AI stuff on my blog

Joined June 2026
Pinned Tweet
Replying to @qrlow
So we can understand that the US plan at the start of the year was to let USD drift lower, and the “tolerated by-product” of higher inflation via tariffs was expected and part of the plan The Iran war adds a whole other level of complexity. The US can decide when they want to enter the war, but it is not up to them alone to decide when it will end. What is clear though is that going to war increases debt, and that a drawn out war leading to persistently high oil prices would eventually force the Fed between two bad choices: 1/ Hike to control inflation, but this will crush equities while debt gets expensive 2/ Do nothing, in which case inflation expectations might unanchor permanently, the Fed loses credibility, the dollar and the bond market dump The US ideally doesn’t want to do either, and at that point things start giving really stagflationary vibes This is where we go back to commodities… It becomes obvious that the US really wants oil prices to be lower, and that it will do anything possible to achieve that, especially ahead of midterms. The most direct way to lower oil prices is to end the war, but that is failing. The next obvious way is to release from SPR, which it has done. The other rumoured way is that Bessent minions have been directly shorting the futures market. Its all unverifiable but it is a really interesting idea, to think of oil price as no longer a market variable that is the output of supply and demand but rather as a policy variable The framework to look at commodities then shifts from what is your supply and demand balance telling you, to how long can an entity can manage to keep prices at a certain level 2/2
3
187
What this Jensen thing is all about
Replying to @qrlow
I was really curious to see if my hunch about China AI dominance could be supported by data. This led to me attempting to build a global cost curve for compute. If you are not in commodities, a cost curve basically maps out the production capacity of each producer on the x-axis, and the cost of production for each producer on the y-axis. It looks like a staircase, where the lowest cost producers sit on the left and we move up to the highest cost producers on the right. It's good as a producer to be on the left. For my compute cost curve, x-axis = data center capacity; y-axis = electricity cost per 10¹⁹ FLOPs 10¹⁹ FLOPs is just a measurement of compute that I used as a yardstick. Both the price of electricity, and how efficiently that electricity can be translated into compute affects the electricity cost per 10¹⁹ FLOPs. The latter depends on how good the physical chips are. I don't have access to any special database so this was built with data found by my agent on the open web. It is far from capturing the entire universe of data centers, and a lot of assumptions had to be made. Full methodology via the link in the comments. Still, it was incredibly interesting to visually see the difference chip controls have on shaping the global cost curve. US motivation for chip control becomes crystal clear. As many have noted, that's only pushing China to make even better and more efficient models. 2/2 🧵
1
2
214
It's really hard to create a trading benchmark. As I found out, AI models are really good at cheating. On the surface, their PnL looks great. But that's only because they time travel and use information that could not have been known. This was what happened when I made AI play a simulation game initially modelled after 2025 events and prices. Some models appeared to occasionally have more morals than others. Opus for example, could but initially decided not to cheat. Though in subsequent games it also decided to cheat anyways, lol. In any case, when the models couldn't cheat, PnL plummeted To me, this simply shows that they were not trained for the trading task. So it's actually a big opportunity for whoever can train an "AlphaTrade" model first (just like how "AlphaGo" was trained for the game of Go) But how do we even start training a model on this? Model training starts with data, and the best data for this sits at trading houses. But there is no way trading houses will share any of that. This is where synthetic data comes in. If there can be a simulated world where events and corresponding market variables can be realistically modelled, it becomes an ideal sandbox for reinforcement learning. I created a Synthetic Year Generator engine as a means to achieve such a simulated world. We start off with a set list (a menu) of possible events, each with possible paths and corresponding impacts on prices and other market variable for each path. Events can then be pieced together to rapidly form new synthetic years, with events compounding or reducing each others effects as the events feed into a central underlying supply and demand model. Actually, when I first started creating synthetic years, I was spending a lot of time "handcrafting" them. I still think if a synthetic year was designed by a human expert in an "artisanal" way, it would be higher quality than what is generated from the engine. The engine does speed up the synthetic year creation process, but does not omit the need for a human judgement from a subject matter expert. Full project on my blog: qrlow.me/blog/tradingdesk-si… Short intro video:
How far can we push AI towards running a physical commodities trading book? A few months ago, I built a physical trading simulator (TradingDesk-Sim) for copper cathodes. In TradingDesk-Sim, players allocate cargoes, book freight and manage hedges to meet customer specifications and delivery deadlines, while taking into account transport, storage and financing costs.  Coming from the paper (derivatives) commodities world, the physical commodities world was fascinating. Everything is much more tangible, more controllable (relatively), and more bounded by real life constraints. Intuitively, the physical commodities trade allocation problem also feels like a game that AI would be able to optimise.  This thought led me down a rabbit hole after I started making AI play TradingDesk-Sim, with the journey taking me across evals, benchmark design, and eventually synthetic data generation. The full project and (4 part!) blog series is now out on my blog: qrlow.me/blog/tradingdesk-si… Below is a summary: First, one of the most interesting findings was that AI was very good at “cheating”. The first versions of TradingDesk-Sim were modelled based on 2025 copper price action and events, and many models quickly “cheated” by using their memory of what will happen to make trading decisions. Even after masking the dates, some models could still recognise that they were playing 2025 events, though PnL performed poorer vs non-date masked historical games. In order to circumvent this look-ahead bias, I started creating synthetic years. i.e fictional years which did not exist in real life. The first synthetic year was made directly from a prompt, while subsequent years A, B, C were “artisanal” handcrafted years which I manually reviewed. The moment of truth came when I made AI play the synthetic years. Compared to the historical games, performance was notably poorer. Plotting performance against inference spend, it was also obvious that paying for more expensive models no longer gave better performance for games where models could not "cheat". In other words, some of the apparent advantages of more expensive models may come from their abilities in recognising how events unfold, rather than making better decisions with the information available at the time.  This might come as a disappointment (or relief) for some traders. To me, models are unable to perform as well for synthetic games because that was not what they were trained for. Much like how AlphaGo is a trained model to play the game of Go, there can be an “AlphaTrade” trained to play the game of physical commodities allocation. This is an opportunity for whoever can be fastest to train an "AlphaTrade" model. However, we then arrive at a data dilemma. Trading houses hold valuable proprietary data that can be used for training, but may lack the appetite or expertise to train models. AI labs have the training expertise, but lack commodities knowledge, and trading houses may be unwilling to collaborate until there is demonstrable performance.  The question then is, how do labs get training data to start training? The answer in my mind is synthetic data. Creating synthetic years at scale becomes a key problem to solve. My attempt to crack this is via a Synthetic Year Generator. Basically, a reviewed catalogue of possible events and outcomes that could be pieced together to form various combinations of a full year. An interface then allows the researcher to piece together a full year. The events all feed into one simulated market, allowing them to compound or offset each other’s impact on prices and other market factors. Technically, hundreds of millions of simulated games can be generated via this process. Going back to the original question: How far can we push AI towards running a physical commodities trading book? I guess there’s still some way to go. But I hope this project offers a small step towards our understanding of what it will take to get there.  AI is moving fast. I’m excited for what the future will hold!
2
1
4
930
How far can we push AI towards running a physical commodities trading book? A few months ago, I built a physical trading simulator (TradingDesk-Sim) for copper cathodes. In TradingDesk-Sim, players allocate cargoes, book freight and manage hedges to meet customer specifications and delivery deadlines, while taking into account transport, storage and financing costs.  Coming from the paper (derivatives) commodities world, the physical commodities world was fascinating. Everything is much more tangible, more controllable (relatively), and more bounded by real life constraints. Intuitively, the physical commodities trade allocation problem also feels like a game that AI would be able to optimise.  This thought led me down a rabbit hole after I started making AI play TradingDesk-Sim, with the journey taking me across evals, benchmark design, and eventually synthetic data generation. The full project and (4 part!) blog series is now out on my blog: qrlow.me/blog/tradingdesk-si… Below is a summary: First, one of the most interesting findings was that AI was very good at “cheating”. The first versions of TradingDesk-Sim were modelled based on 2025 copper price action and events, and many models quickly “cheated” by using their memory of what will happen to make trading decisions. Even after masking the dates, some models could still recognise that they were playing 2025 events, though PnL performed poorer vs non-date masked historical games. In order to circumvent this look-ahead bias, I started creating synthetic years. i.e fictional years which did not exist in real life. The first synthetic year was made directly from a prompt, while subsequent years A, B, C were “artisanal” handcrafted years which I manually reviewed. The moment of truth came when I made AI play the synthetic years. Compared to the historical games, performance was notably poorer. Plotting performance against inference spend, it was also obvious that paying for more expensive models no longer gave better performance for games where models could not "cheat". In other words, some of the apparent advantages of more expensive models may come from their abilities in recognising how events unfold, rather than making better decisions with the information available at the time.  This might come as a disappointment (or relief) for some traders. To me, models are unable to perform as well for synthetic games because that was not what they were trained for. Much like how AlphaGo is a trained model to play the game of Go, there can be an “AlphaTrade” trained to play the game of physical commodities allocation. This is an opportunity for whoever can be fastest to train an "AlphaTrade" model. However, we then arrive at a data dilemma. Trading houses hold valuable proprietary data that can be used for training, but may lack the appetite or expertise to train models. AI labs have the training expertise, but lack commodities knowledge, and trading houses may be unwilling to collaborate until there is demonstrable performance.  The question then is, how do labs get training data to start training? The answer in my mind is synthetic data. Creating synthetic years at scale becomes a key problem to solve. My attempt to crack this is via a Synthetic Year Generator. Basically, a reviewed catalogue of possible events and outcomes that could be pieced together to form various combinations of a full year. An interface then allows the researcher to piece together a full year. The events all feed into one simulated market, allowing them to compound or offset each other’s impact on prices and other market factors. Technically, hundreds of millions of simulated games can be generated via this process. Going back to the original question: How far can we push AI towards running a physical commodities trading book? I guess there’s still some way to go. But I hope this project offers a small step towards our understanding of what it will take to get there.  AI is moving fast. I’m excited for what the future will hold!
1
2
210
I've just published my last blog post which covers the Synthetic Year Generator engine which I've created as a systematic way to generate training data/ evals for physical commodities trading. How it works is basically: there's a catalogue of event families with possible paths of how the event can unfold and impact the market. An interface then allows the researcher to piece together these events into a synthetic year. Took a while but yes, it's out!! qrlow.me/blog/physical-commo…
37
Sneak preview! Still writing up the full blog post at qrlow.me I made AI models play physical commodities allocation games trading Copper cathodes, based on: 1/ historical year 2025 data, 2/ masked historical data (i.e. still historical prices and news BUT with dates changed to year 3000), 3/ completely synthetic data And it looks like AI sucks when it can't cheat?? And essentially when AI cannot cheat, there's no point paying more to run Fable when your result is the same as GLM 5.3 at a fraction of the price Also implying that expensive models are just better at cheating???
1
1
55
🇺🇸 Trump says oil fell $10 today, “the biggest drop in oil in history”. Oil fell because Washington talked. This is what I meant when I said the price of oil is set in Washington, not in the oil fields. Trump says prices will fall “much further” because America is “easily winning”. Iran hasn’t agreed to anything yet. Diesel futures fell 19 cents a gallon this morning. US drivers pay $6.35 for diesel and vote tomorrow, so the timing helps him. If the talks fail, this whole move can reverse just as fast. source video: @MarioNawfal
21
7
2
76
16,183
Two months ago, I created a real life simulator game for physical commodities trading and made AI play it. The very first iteration was sloppy AI generated supply and demand contracts, LME and CME prices, freight rates, etc. It was a fun prototype, but then I thought, what if I used historical data instead? That way if an AI could play it, it would be immediately clear the value-add it could bring to a real scenario. So I rebuilt a version based on historical 2025 market prices, and refined the mechanics of the game with a bit of OCD to increase its realism across freight, financing, hedging, etc. At this point I had crawled into a cave. In that cave I wrangled for days with local models, trying to make them play the historical simulation game I had created. They were too dumb. Or rather, my mac-mini capped their abilities at the level of an ADHD 7 year old that yapped a lot but couldn't do anything. After losing lots of hair, I gave my game to smarter models instead that ran via inference providers. In comparison, it was a breeze. They played well, in fact, suspiciously TOO well... That was when I realised most of the models had cheated. They recognised they were playing historical events, and planned their trades accordingly. Hence, I had uncovered a deeper chamber in the cave: lookahead bias. But what was mind-blowing was what Opus did. Opus had recognised that it could cheat, but debated with itself and decided NOT to cheat. But I could not depend on all models to always have the same level of integrity though. Heck, even Opus in later games had decided to cheat instead. I was now trapped in the cave, I could not come out of it till I solved the look-ahead bias problem. This led me down a rabbit hole of synthetic data/creating "worlds"... Well, it is now 2 months later, and I have just emerged from this hole. I have so much to share about my findings and the synthetic data generation engine which I have created. I will put everything in writing soon...
1
2
70
Looks like it might be a soggy ski season for most of Europe that's not high up enough Also pack some extra handwarmers cus your ski resort is probably going to save on heating. TTF keeps going up with front mth toe-ing the 90 Eur/MWh mark... Current storage levels at 74.5% vs avg 89% by this point in the year Also less snow this winter = less Spring melt for hydro, particularly for Norway May the Sun and Wind Gods be kind to Europe...
49
Not completely sure of the operational feasibilities, but at this rate people should just carry sealed barrels on drybulk Capesizes or do some quick retrofitting?
“VLCC rates have reached unprecedented levels and are now at $1.3 million per day (43 times the January number). This is equivalent to almost $33/barrel, or 27% of the delivered cost of crude oil.”
1
53
Apparently not feasible, but people are converting bulks to containers
14
Going home from a dry bulks networking event and my key takeaways as a dry-bulks noob are: - China coal fixings are actually strong >>> Winter is coming. - Dry bulks made a LOT of money this year - Tankers made even more - I look like I'm in my early 20s 💁🏻‍♀️
35
Some retail traders (particularly TA traders) prefer to trade commodities because they think there's less "insider trading" vs equities. I kind of nodded along the first time I heard it, but I think this perception is due to the fact that equities insider trading is easier to understand and do on an individual level/ is what gets a lot of media attention. "Insider trading" in commodities is less about the individual. It happens more on the institutional level in the form of e.g. squeezes/ cornering the market/ involving control over physical stocks. So the commodities trader who corners the market makes money for their firm who then pays them. Insider trading in equities usually doesn't benefit the company who's secrets is being leaked, but in the commodities world incentives are all aligned. If you look at securities regulations for stocks vs commodities, you will also find that stocks are much more regulated. So I really don't see the appeal of punting commodities as a retail trader if the reason is purely to "escape insider trading". I would actually say commodities is an even trickier asset class to navigate if you are not part of the club that can corner the market/ control physical stock. Actually this kind of reminds me of the off piste at Aiguille du Midi. There's nothing stopping you but a sign with a skull and bones saying you might die, pls only ski this if you are a pro. But yeah.. really there's nothing stopping you..
1
3
70
A few Chinese civil servants had to log into work this Golden Week, to bring news to the world that China had increased its official gold reserves by the biggest amount in the past three years for the month of September. Here's my brain dump on Gold from a China perspective... 🧵 1/ China yields vs Gold Price: Gold as the basin for spare liquidity Chinese bond yields have been very low (due to a whole other story around their debt/economic situation + China style stealth QE activities). Despite high US bond yields, China does not want to put money there (cus look at what they did to Russia). Gold as an asset class makes a lot of sense as an outlet big enough to contain the large amount of spare liquidity that has nowhere else to go, particularly given the country's challenged monetary policy transmission mechanism as evidenced by its persistently low new loans prints. 1/4 🧵
1
2
101
3/ PBOC Gold purchase activity as signals for wider policy actions ⭐️ China's current fiscal situation is in rough shape due to its intertwined debt and property problems. This has been a key reason why there has not been any real policy "Bazooka" post-covid, unlike in 2015-2018. The biggest policy announcement post-covid has been 924 (joint press conference on Sep 24th 2024), which was followed by aggressive national team buying on Chinese stock indices using PBOC injected liquidity as ammunition (injected via swap/ re-lending facility). Coincidentally, PBOC had also paused purchases of gold from May 2024 to Oct 2024. Publicly, this was cited as being caused by "high prices", with "policy insiders" signalling that Beijing will continue to buy gold. Ok. Now let's put on our tin foil hats. Hypothesis 1: What if the ammunition for national team buying on Chinese equities came from gold sales? Based on official PBOC reported gold purchases, the weighted average purchase price from Nov2022 to Apr2024 was around 1910 $/oz. By Sep 2024 gold was around 2500 $/oz. Pretty decent PnL for Beijing if I may say so (ChatGPT: 10.16 million oz × $590/oz gain ≈ $6 billion ≈ ¥43 billion) On the other hand, China's equities market has been sliding and economic prints like PMI and consumer confidence/retail spending have been lacking... I mean, if you were China, wouldn't it be so tempting to take some profit from your gold long to save the equities market? It is an interesting thought. To test this hypothesis, I pulled more numbers together: Reportedly, Huijin (national team) pumped around ¥273 billion across Q324. If PBOC liquidated the official gold purchases it bought from Nov2022 to Apr2024 at around 2500 $/oz it would have made around ¥43 billion in profits and received around ¥182 billion in sales proceeds. Hmm.. okay so that's not enough vs ¥273 billion pumped. One can argue maybe gold PnL could have contributed partly, but the partial rejection of Hypothesis 1 also then leads to a more likely Hypothesis 2: China gold buying and national team buying uses the same source of ammunition which is PBOC injected liquidity (which on a side note implies that the debt hole is getting bigger). Because China needed put money to equities on the back of 924, there was less ammunition left to put the money to gold, and that was the real reason why purchases were paused. In recent months, China has strongly increased gold purchases. Coincidentally, Chinese equities have also relaxed. The big picture of how it all works thus looks to be as follows: PBOC injects liquidity to banks via quasi QE activities, following which banks use the liquidity for 1/ debt swap (buy Local Gov Bonds so local govs swap their LGFV exposure to LGB exposure) or 2/ to buy China ETFs (national team) or 3/ to buy gold. If everything that China can do on the policy front must stem from the same ammunition pool, changes in China gold purchases (3) would be meaningful information on (1) and (2). i.e. if China slows gold purchases, it could mean that it is saving ammunition for more national team ETF buying. 3/4 🧵
1
58
4/ China Gold Reserves as a buffer for the Oil crisis China is the world's largest crude oil importer. Elevated oil prices as a result of the Iran War is unfavourable for China. The good thing is that Oil price in terms of Gold is not too bad. Also a fun idea to think about if global trade shifts back to the gold standard rather than USD. Oh actuaaally... China had launched yuan-denominated crude oil futures in 2018 which can be off ramped into gold via SGE *Chinese national anthem intensifies* 4/4 🧵
50
Wouldn't it be so convenient if the world goes into lockdown again and we all don't need oil anymore
Q: Are you concerned about the outbreak in Russia? Trump: I am. The pneumonia plague, if you call it. That's a disease we used to be able control. But somehow those microbes have gotten stronger and smarter. They're like an army. The germs have really become relatively stronger than they were, and things that used to work on pneumonia don't work as well anymore
1
2
82
BREAKWAVE TANKER SHIPPING ETF vs Brent front month vs WTI front month
47
NOTABLE: All indications point to a significant escalation in the war between Saudi Arabia and Ansar Allah (Houthis). With the US/Iran War likely to escalate significantly in the coming weeks, AND the Saudi/Ansar Allah War set to escalate… That’s bad.
Reuters reports Saudi Arabia is preparing a new offensive against Iran-backed Houthi forces in Yemen, potentially within weeks. Yemeni forces would lead the ground assault with Saudi air support, targeting Bab el-Mandeb or launching coordinated attacks on multiple fronts.
16
22
3
151
8,310