Recent
- Research noteBT-2026-0235
The last-half-hour reversal survived out of sample
A pre-registered test finds the end-of-day reversal in S&P 500 stocks still near 3 basis points a day, but weaker since 2020.
2 minSource: arXiv
- Research noteBT-2026-0234
A paper says price-volume correlations are always zero
A new paper argues the standard empirical price-volume correlation is zero by construction, and proposes measuring against squared volumes.
2 minSource: arXiv
- Research noteBT-2026-0233
A Sharpe near 3 that fell to 0.26 on the holdout
An open-source LLM research loop for trading is built so the model proposes and only trusted code decides. Its best lesson was a failure.
2 minSource: arXiv
- Research noteBT-2026-0232
Round numbers hold more orders, not the price
A year of order book data for five Nasdaq stocks finds more resting liquidity at round prices but no extra stickiness.
2 minSource: arXiv
- Research noteBT-2026-0231
You take the dividend drop without the dividend
Premium-index funding on crypto equity perpetuals cannot pay a dividend, and one venue's design moves it to another.
2 minSource: arXiv
- Research noteBT-2026-0230
Six defects none of the summary numbers showed
A live intraday equity system was compared against its backtest replay; the drift never appeared in the statistics.
2 minSource: arXiv
- News briefBT-2026-0229
Binance routes thin local-currency trades via deeper books
Auto Order Routing compares a local pair with a more liquid intermediary path and sends the order to whichever quote is better.
2 minSource: Binance
- Research noteBT-2026-0228
Execution agents that learn to punish
Two reinforcement learners in a liquidation game beat the Nash benchmark, then retaliated against a deviation at no cost to themselves.
2 minSource: arXiv
- Research noteBT-2026-0227
The architecture is not what is unclear
An audit of 30 studies finds agent design reported more clearly than the execution assumptions needed to judge whether a trading result means anything.
3 minSource: arXiv
- Research noteBT-2026-0226
A ranking loss and a profit threshold
A convolutional model scores 0.6316 on area under the curve for a seven-day Bitcoin move — and the paper is more interesting for how it is judged.
2 minSource: arXiv
- Research noteBT-2026-0225
When going first is worth something
Three authors work out when a constant-function market maker is at rest, and when the order of two traders decides who gets the better fill.
2 minSource: arXiv
- Research noteBT-2026-0224
When to move the range, and what it costs
Five authors model concentrated liquidity provision as impulse control and learn the rebalancing policy, reporting a compressed lower tail rather than a higher mean.
2 minSource: arXiv
- Research noteBT-2026-0223
Quoting at several levels at once
A reinforcement learning market maker places orders across the book rather than at the touch, and is tested against three kinds of counterparty.
2 minSource: arXiv
- Research noteBT-2026-0222
An agent that reads the chain and the news
CryptoTrade pairs on-chain metrics with off-chain signals and reflects on its own past trades before deciding the next one.
2 minSource: arXiv
- Research noteBT-2026-0221
Two in a thousand graduate, eight times more with a channel
A survival analysis of 832,941 token launches puts a number on something the market treats as folklore.
2 minSource: arXiv
- Research noteBT-2026-0220
Generating order books in ten solver steps
A flow-matching model produces limit order book trajectories at a fraction of the sampling cost of diffusion, and transfers to instruments it never saw.
2 minSource: arXiv
- BenchmarkBT-2026-0219
Language models are poor traders and useful analysts
Given execution authority they underperform. Given a research task they save hours. The distinction is not subtle.
SpansAISECFIN
10 min
- AnalysisBT-2026-0212
The ordering supply chain professionalised
Extraction moved from opportunistic scripts to an industry with tiers and contracts.
SpansAISECFIN
7 min
- Research noteBT-2026-0205
Backtests on model-driven strategies leak more than usual
The model was trained on the period you are testing. That is not a detail.
SpansAISECFIN
6 min
- Field reportBT-2026-0198
Latency arbitrage moved to inference time
When two desks run similar models, the edge is who answers first.
SpansAISECFIN
8 min