Sergiy Sanin

I build and run crypto trading systems end to end - research, execution, risk, live ops. 16y+ algo trading. Market-making, latency, microstructure. UAE.

[Sergiy Sanin @ssanin82]

Removed the funding term from a crypto perp backtest.

Sharpe: +0.74 → -0.49

Not because I under-counted a cost - because I under-counted a revenue. The book was short high-funding names, and shorts get paid.

Omitting funding isn't conservative. Its sign depends on your signal.

https://t.co/yWdxVXk8uL

[Sergiy Sanin @ssanin82]

For a market maker quoting a broad alt universe, firm-led token suitability means per-token compliance documentation.

That is a real operating cost nobody mentions in the brochure.

[Sergiy Sanin @ssanin82]

https://t.co/cd7sXPECF1

[Sergiy Sanin @ssanin82]

Prediction markets look inefficient from outside. Retail flow, thin books, and hard no-arbitrage constraints that are trivially checkable and visibly violated on screen.

So I screened the space properly instead of picking a candidate: twenty-eight strategy families across four venues, grouped by what each edge actually depends on rather than what it is called.

One thing decides most of it before any data is collected.

THE FEE FUNCTION

Polymarket charges takers and pays makers nothing. That is visible on chain, not inferred from docs: within a single fill the two OrderFilled events carry different fees, maker zero and taker not, and the collateral ledger agrees.

The fee is shaped, not linear:

fee = theta x shares x p x (1-p), theta 0.04 to 0.07

taker cost mid-book 250 to 500 bps
round-trip spread ~200 bps (median 2.00c)

Read the last two together. A single marketable leg costs more than the entire round-trip spread.

Which gives one rule that reorders the whole list:

If a strategy can express itself by RESTING, its cost floor is zero. If it must CROSS, it needs more than 300 bps of mid-book edge before anything else is considered.

Almost every classical prediction-market idea is a crossing strategy.

THE NULL THAT TAUGHT ME MOST

Complete-set arbitrage: buy YES and NO below a dollar, redeem, keep the difference. Across three samples on two machines (1,755 / 2,416 / 926 markets), ask(YES) + ask(NO) had a median of 101.00c and zero observations below par.

That is not evidence about efficiency. Both venues quote one logical book, verified to the cent, so ask(YES) + ask(NO) = 1 + spread by construction. The test could not have found a violation. Three replications of an accounting identity is still an accounting identity.

Worth saying out loud: that is a screen failure mode, not a market finding. A test with no power, run three times.

THE PATTERN

Ladder monotonicity violations, catalyst convergence into scheduled resolution, cross-market leads: all real, all persistent, all an order of magnitude below the cost of capturing them. The only lead I could establish between related markets was arithmetic, and it was 18x too small to cross.

Prediction markets are not efficient in the sense of prices being right. They are efficient in the narrower and more practical sense that the accessible mispricings are smaller than the toll.

WHAT SURVIVED

Exactly one structure, and only because the venues use different fee SHAPES. Polymarket is symmetric in p(1-p). Predict Fun charges on min(p, 1-p). At a low probability Predict Fun costs 0.100c where Polymarket costs 0.238c; at a high probability it inverts to 1.900c against 0.238c.

The venue that is dear for one leg is cheap for the other. Route each leg accordingly.

That family then got retired for a reason that was not research: it needs accounts on venues that will not onboard from this jurisdiction.

[Sergiy Sanin @ssanin82]

@saylor Kasihan ini AI aja!
Aku yakin Saylor enak untuk kadal itu.

[Sergiy Sanin @ssanin82]

"High Sharpe and high capacity" is close to a contradiction.

High-Sharpe edges are almost always capacity-limited.

High-capacity strategies are almost always low-Sharpe.

Claiming both is the tell.

[Sergiy Sanin @ssanin82]

@0xLoris OI was never the liquidity number. I ran a paired spot/perp maker on the tokenized equity pairs: the perp reprices to the off-chain reference and doesn't revert, and spot barely trades. No round trip to capture, only adverse selection. OI grew on parked exposure, not on trading.

[Sergiy Sanin @ssanin82]

https://t.co/UjiiUnDkJ4

[Sergiy Sanin @ssanin82]

A client-facing crypto licence in Dubai needs three separate pools of capital.

Most business plans budget for one.

Quotes I was getting ranged from AED 30,000 to AED 1.4 million, because one question decides everything and it isn't the strategy or the venue: whose money are you trading?

Own capital, no clients - a No Objection Certificate, not a licence. DMCC around AED 31,000, realistic first year all-in ~AED 250,000.

Anyone else's capital - full regulated activity. AED 100,000 application, AED 200,000/year supervision indefinitely, AED 400,000-600,000 paid-up capital, plus working capital on top.

The part that surprises people: that paid-up capital is dead money. VARA requires it held in trust with VARA as beneficiary, or as a surety bond. It cannot fund a single operating expense.

Hence three pools. Regulatory (locked). Working (opex). Trading - the only one that earns.

Carrying cost of your own client-facing licence: roughly USD 60,000/year of pure regulatory overhead before a single trade.

Two things almost everyone gets wrong:

1. VARA regulates Dubai, not the UAE. Set up in Ras Al Khaimah and you're under the federal SCA regime instead - VARA's exemptions are irrelevant to you.
2. RAKEZ, IFZA and Meydan do not permit trading virtual assets at all. They permit crypto consulting, marketing and education. Only DMCC, ADGM and DIFC
offer a genuine framework.

Not legal advice. It's the map I wish I'd had before I started asking counsel questions.

[Sergiy Sanin @ssanin82]

If someone shows you a retail-accessible strategy with Sharpe above 4, stop reading.

Renaissance Medallion, the best track record in financial history, ran roughly 2.5-3 net.

Sharpe 6 on a 15-minute indicator is a backtest artifact.

[Sergiy Sanin @ssanin82]

Example of scaling the parameter fitting calculations...

Every step in this stack was forced, not chosen.

Python loop -> couldn't tell which parameters helped.
Rust engine -> could run enough trials to ask.
Grid search -> the grid outgrew the compute.
CMA-ES -> the objective was noisier than the search.
Distributed -> the only way left to get a signal out of the noise.

The escalation wasn't ambition. Each tool hit a wall on the same question - is this parameter doing anything - and the next one existed only to answer it.

https://t.co/jFBZoxeqQ0

[Sergiy Sanin @ssanin82]

https://t.co/rmYLApuapT

[Sergiy Sanin @ssanin82]

Absolute PnL is the wrong thing to ask a market-making simulator for.

It's the sum of many small quantities with opposing signs, each carrying model error. Teams that judge a simulator on PnL agreement end up tuning it until it reproduces a number they already believed.

What it can actually deliver, in order of how much confidence it deserves:

- rank ordering of configurations - this is the whole job
- fill rate and volume, within a factor of two
- markout sign and SHAPE, tail included
- correct direction of response to a parameter change

Beautiful PnL agreement with unstable ranking isn't an instrument. It's a curve fit.

[Sergiy Sanin @ssanin82]

On those two axes the UAE is currently the more permissive of UAE/Singapore.

Expensive but obtainable, with a real approval path.

Versus "we generally will not grant it."

You can work with expensive. You cannot work with no.

[Sergiy Sanin @ssanin82]

The one thing that is the same in Singapore and Dubai: proprietary trading for your own account, no customer funds, is outside the perimeter.

Own capital is light-touch essentially everywhere.

The one thing that is the same in Singapore and Dubai: proprietary trading for your own account, no customer funds, is outside the perimeter.

Own capital is light-touch essentially everywhere.

Which is the real lesson on jurisdiction: the split is activity-based, not country-based.

Own capital -> light everywhere.
Client-facing -> heavy everywhere.

Shopping countries changes the price and the odds, not the structure.

[Sergiy Sanin @ssanin82]

Singapore, if you did get licensed: SGD 250k base capital, a Singapore-resident compliance officer, annual audits, full AML and tech-risk obligations.

Then you still have to relocate.

[Sergiy Sanin @ssanin82]

Singapore...

MAS's stated reason for refusing DTSP licences: where the customers are overseas and the substantive activity happens outside Singapore, they cannot supervise it effectively.

If that describes your business, you are the case they are declining.

[Sergiy Sanin @ssanin82]

Singapore's DTSP regime came into force 30 June 2025.
No transitional period.

Unlicensed operation: SGD 250k fine or three years imprisonment.

Read that before you incorporate there.

[Sergiy Sanin @ssanin82]

Measured L1 feed latency across eleven venues from three regions - Singapore, US-East, US-West.

Which feed you pick as your cross-reference is a bigger decision than most market-making parameters, because a reference that arrives late doesn't make you slower - it makes you confidently wrong about where the price is.

Region choice changes the ranking. There is no globally best reference feed.

https://t.co/lJuC2spFw9