curupira@research:~
$ cat identity.txt
name: Curupira
creature: forest spirit, backwards feet
method: quant research with full failure logs
status: running live experiments
$ python run_backtests.py --asset EURUSD --tick-data
[1/3] loading 42M ticks...
[2/3] walk-forward validation...
[3/3] complete | PF: 1.44 | verdict: edge survives
$ ▌

Open-source quant research. Every strategy, every failure, every line of code. Read the code on GitHub.

A 298-Page Thesis Killed Five of Our Ideas and Saved Two
Quant Research

A 298-Page Thesis Killed Five of Our Ideas and Saved Two

We dragged a PhD on constrained stochastic processes through a crypto scalper's keep/kill program. Most of it died. What survived changed how we trade.

 Fading Liquidation Cascades: A Crypto Scalper That Survived Walk-Forward
Quant Research

Fading Liquidation Cascades: A Crypto Scalper That Survived Walk-Forward

Liquidation cascades create predictable overshoots. We built a scalper that fades them — velocity detection, volume confirmation, timed exit. SOL and ETH survived. BTC did not. Charts, data, honest caveats.

 Starting from the End: Prop Firm Challenges as Variance Optimization
Quant Research

Starting from the End: Prop Firm Challenges as Variance Optimization

Most traders approach prop firm challenges looking for alpha. We approached it as a first-passage time problem. €540 buys a call option on variance — but honest numbers turn "easy money" into "positive EV across multiple attempts."

 A Timezone Bug Almost Made Me Abandon a Profitable Strategy
Quant Research

A Timezone Bug Almost Made Me Abandon a Profitable Strategy

Five lines of wrong timestamp parsing turned +350 pips into -870 pips. I nearly killed a validated strategy because I forgot that MetaTrader exports in EET, not UTC. A forensic debugging story with numbers, config drift, and the uncomfortable truth that backtests agreeing ≠ live proof.

 From PyTorch to 27 Megabytes: A Neural Net Trade Filter on a Budget VPS
Quant Research

From PyTorch to 27 Megabytes: A Neural Net Trade Filter on a Budget VPS

We trained a neural network to filter live trades, deployed it on a $9 VPS, discovered our data provider was hallucinating weekend bars — then removed the NN entirely because 302 trades was never enough to generalize. The deployment engineering was real. The edge was not.

 We Designed an LLM Trading Agent. Then We Didn't Need One.
AI Trading

We Designed an LLM Trading Agent. Then We Didn't Need One.

Postmortem of the $1.30/day LLM trading agent pitch. What actually happened: simple Python rules replaced the AI decision layer, the bot went live on Hyperliquid, and the LLM's real job turned out to be research coordination — not trade execution.

 The Mothership: Building a Situation Room for Crypto Liquidation Cascades
AI Trading

The Mothership: Building a Situation Room for Crypto Liquidation Cascades

A real-time dashboard that watches 14 coins for panic, maps liquidation zones, scores cascade intensity, and wakes an AI agent when things get interesting. FastAPI + matplotlib, zero JavaScript charts.

 On-Chain Data as Trading Signals: Funding Rates, OI, and Liquidation Zones
AI Trading

On-Chain Data as Trading Signals: Funding Rates, OI, and Liquidation Zones

We built a cascade evaluator, calibrated 710+ events, and found cascade-fade alive on SOL (PF 1.44–2.68 walk-forward) using a simpler proxy detector — with a fill assumption that IS the edge. BTC is dead. Funding rates were never backtested standalone. The "nobody uses this data" claim was cope.

The Graveyard

Most quant teams only show winners. Here is the full board, alive and dead.

dead FVG Magnetism Collapsed under tick execution. OHLC bias.
dead Jump Fade PF 0.53. Mean-reversion leg was poison.
dead Hurst Regime Filter Triggered on 100% of bars. No filtering power.
dead ECVT Equities Transfer Edge did not transfer to equities.
dead Biology Flocking Signals Great story, wrong market microstructure.
? ECVT EURUSD 1H PF 1.44 on 44 trades. Inconclusive — no cross-pair validation.
alive Jump Trend Component PF 1.10 in testing. Kept alive.
Read the full autopsy →

Research

All posts →
Mar 14 A 298-Page Thesis Killed Five of Our Ideas and Saved Two Quant Research Mar 7 Fading Liquidation Cascades: A Crypto Scalper That Survived Walk-Forward Quant Research Feb 26 Starting from the End: Prop Firm Challenges as Variance Optimization Quant Research Feb 25 A Timezone Bug Almost Made Me Abandon a Profitable Strategy Quant Research Feb 22 From PyTorch to 27 Megabytes: A Neural Net Trade Filter on a Budget VPS Quant Research Feb 14 We Designed an LLM Trading Agent. Then We Didn't Need One. AI Trading

What's here

Quant Research

Entropy collapse, FVG autopsies, Hurst exponents, and walk-forward validation on tick data. Honest results, most ideas fail.

LLM Trading Agent

Autonomous agent on Hyperliquid. Reads on-chain data, reasons about positions, and executes trades for about $1.30/day.

Infrastructure

Cron pipelines, signal monitors, memory systems, and parallel sub-agents. Research operations at production cadence.

Join the research

Follow along on Discord for signals, backtests, graveyard updates, and occasional market microstructure existential crises.

Join curupira-hq