Electronic Trading Architect | Strategic Advisor

If you collect your trading logs, you're sitting on a gold mine. Hidden in logs FIX/JSON/CSV: - the fills losing pnl in the first second after execution - the cancels losing queue position to internalization - the venues "quietly" adverse-selecting your side - the reject paths running 10x slower than the happy path - the adverse-selection buckets eating most of your alpha The signal is right there in the logs. You just need to know what to read and how. We've found an accurate, systematic way to extract every leak. With a fully deterministic engine, no LLM, no inference, just the math. Works for #FX, #crypto, #equities, anything with execution logs.

#meme of the week #trading #markets

Updated the profile photo. Still debugging markets. Now in higher resolution 🤣

Routing audit gap from today's thread: 10-20 bps spread between fragmented and optimized routing. At $10B notional, midpoint = $15M/yr. Latency gets instrumented to microseconds. Routing logic gets a quarterly review. Thread: x.com/sisSoftware/st…

The desk that can answer those 4 in basis points — not milliseconds — already knows where it sits. Most can't. If your desk can, the venue that surprised you is the comment worth reading. hftadvisory.com/microstructure… #executionquality #algotrading

Minimal routing audit checklist: → Segment fills by venue. → Arrival-price slippage per venue, not blended. → Compare vs. published make/take + actual fill rate. → Has routing logic been backtested against the last 90 days of venue microstructure shifts?

Most desks have audited their latency down to microseconds. Very few have audited their routing quality in basis points. Those two numbers are not correlated. And one of them moves more P&L. #hft #electronictrading

@KaV_2599 hardware isn't the hard part, you can buy that. the hard part is getting position state, fills, and risk all wired together at desk latency. that takes years.

Upfront cost is more than training the largest models. I'm talking about 10s of billion of dollars of Hardware, from co-location data centers, to semi custom FPGAs to multimillion dollar oscilloscopes. Their hardware research lab is probably as advanced as that of a semiconductor company plus quant hiring is expensive in general.

@ArdiNSC seen this a few times. when CVD keeps climbing and price barely moves, the ask is usually a POV algo with a capped rate. how fast the same level reloads after each lift is the tell.

Someone is sitting in the bitcoin:native limit order book and distributing directly into the bullish spot flow. During the rally up to $79.5K, Spot CVD was making fresh highs, and we saw price begin to break out directly because of that. Now that parallel relationship has diverged. Spot CVD is even higher than it was at the top of the range, but price is gradually moving lower. All we're seeing here is the spot bid getting absorbed before it can move price, meaning there are relatively large players with size sitting behind the books and selling into that strength. Not surprising in the least. These environments are specifically manufactured to get people bullish again, so there is enough demand to distribute supply into. Now it's just a matter of patiently waiting to see when Spot CVD begins breaking down.. 👁️

@Maverick_Quant well, as I mentioned. If you can tap into their level 3 market data, then you can. However, in defi and web3 is hard to get

Impossible to know - trade to cancel ratio and cancel orders ! Exchange knows this but cannot deduce from Orderbook

@ts_crpt PvP runs at whatever latency the venue exposes. tradfi HFT in microseconds, crypto CEX microseconds internally but ms (or even more) to clients, L1 DEX hundreds of ms. adversary mix shifts more than the timescale.

@sisSoftware 結局のところPvPだということですね。 具体的にはどのぐらい時間軸なんですか？tradifiとブロックチェーンの時間のオーダーは桁が違うとは思いますが。

Jane Streetはじめ大手HFは従来のlatency重視のHFTからMLによるmid termの予測を使ったHFTに移行しているとのこと。 これはハイリキの上位アカウントの挙動と合致するように見える。 app.hyperliquid.xyz/trade/HYPE/USD… 例えばこのユーザーは、入出金のTXからおそらく取引所横断型のロジック"でない"と推定されるユーザーだが、mid priceとオーダーの距離の変動の挙動から30分ごとに目標を変更していると推定される。 これまでの方針として、週次や月次などの長期の予測の戦略が運用資産額のスケール可能性が高いとしてきた。 一般的な執行コストで考えるとその仮説は正しいが、収益額でトップのHFがHFTであるという事実からHFTはスケールが可能であるどころか、収益に関しては最もスケールすることが示されている。

@bwein_ the leak is in placement before cutoff. faster book reads forecast the clearing price better, so you submit orders positive-EV under that forecast. uniform clearance fixes price + fill direction tracks forecast quality.

@sisSoftware thanks for the reply ser. would you plz elaborate on how predicting state before cutoff, particularly in the context of a batch auction with uniform clearance, still leaks the same amount of value?

All this for fractions of a second. I believe the market is more efficient without the HFT latency arms race. To date only partial solutions exist. Deliberately synchronizing order matching, strategy execution and information offers a complete solution.

Cancel inference from L2 snapshots: 30-50% degradation vs. true L3 in busy books. No paper publishes that number. Practitioners know it. It's why the architectural decision between L2-priced and L3-priced execution is bigger than the data fee. #marketmicrostructure #hft

Test: cancel-OFI alone, trade-OFI alone, combined - same instrument. If combined doesn't dominate, two hypotheses: snapshot rate aliases cancels, or latency exceeds signal half-life. Your venue and the break point is the comment I want to read. electronictradinghub.com/?p=2078

On half-life: the Cont signal runs 10-30 seconds in liquid equities. A 10-second half-life and a 12-second data pipeline are not compatible. The architectural fix: decompose OFI by event type, and prioritize L3 venues where signal quality matters.

97% of limit orders cancel before they ever trade. The dominant event in any order book is not a fill. It's a cancellation. If your signal stack consumes the trade tape and discards the cancel stream, you model the residual. A breakdown: #hft #marketmicrostructure