Why We Open-Source Our Quant Research

The Graveyard Is the Product

We’ve tested 38 strategies. 36 are dead.

Not “underperforming.” Not “needs more data.” Not shelved for later review. Dead. Killed with tick data, walk-forward validation, and the kind of honesty that makes you rm -rf code you spent three weeks writing at 2am because the numbers don’t lie, even when you wish they did.

The two survivors: IPDA CE — live, PF 1.21, thin edge but real, 6 live trades tracking within margin of backtest. And Cascade-Fade Scalper on SOL — PF 2.56 on paper, but fill-dependent, meaning it might evaporate the moment it touches a real order book with actual slippage. We’ll see.

Hit rate: about 5%. That’s the number nobody publishes. For every strategy that survives contact with real data, nineteen don’t. We publish all twenty.

The Kill List

I keep the list because the breadth matters — we didn’t just try one thing and give up.

Entropy and information theory. ECVT — entropy collapse on EURUSD hourly. Our first real signal. We were proud of this one. Called it “PF 1.44, inconclusive” early on, kept digging. Pair-specific (EURGBP: -63.6 bps), OHLC-specific (ticks destroyed it), bombed on equities (SPY: 9 trades, -88 bps, 89% stop-outs). A signal that only works on one pair at one timeframe in one bar resolution isn’t a signal. It’s a coincidence wearing a lab coat.

Price action. FVG magnetism, gap-fill physics, ICT concepts — the sacred cows of retail trading Twitter. OHLC daily inflated the profit factor by 4×. On 42 million ticks: dead at every timeframe. Monotonic degradation from daily to 5-minute. The autopsy is one of our most-read posts, probably because it takes everyone’s favorite narrative out behind the shed with actual data.

Regime detection. Hurst exponents, DFA cross-validation, regime-adaptive switching. On crypto hourly: 100% of bars classified as “trending.” Every single one. Zero information content. On forex daily: a faint signal that degraded when combined with anything useful. Like catching a liar — the exponent tells you something is there, then crumbles the moment you try to act on it.

Chaos theory. Lyapunov exponents, fractal dimension, attractor reconstruction. Gorgeous mathematics. Zero predictive power at tradeable horizons. The kind of thing that wins you a PhD and loses you money.

Biology-inspired. Flocking models, predator-prey dynamics, swarm intelligence applied to order flow. Wrong asset class, wrong timescale, wrong everything. Starlings aren’t limit orders. Nature metaphors don’t survive contact with microstructure noise.

Network topology. Graph-theoretic approaches to correlation regimes, centrality-based pair selection. Interesting for portfolio construction. Useless for generating a signal that says “enter here, exit there.”

Performative market making. Crowding detection via funding rates, epsilon estimation from public data. The most beautiful thesis we’ve explored. The data to test it doesn’t exist — funding rates update every 8 hours versus 5-minute bars (96:1 mismatch), no free historical LOB. Gorgeous on a whiteboard. Unmeasurable without paid data feeds. Theory ≠ tradeable, no matter how elegant the theory.

Every single one of these looked promising in the first backtest. Every. Single. One. Then we tightened the screws — tick data, walk-forward, out-of-sample, cross-asset — and they broke.

Why Publish the Corpses

This is the question I get asked like it’s some gotcha. “Why give away your research?” As if the research itself is the moat.

Each dead strategy postmortem takes hours to write. The methodology, the code, the specific fracture point, the lesson buried in the wreckage. Nobody pays us. Nobody’s hiring based on kill count.

We do it because of what it does to us.

You can’t lie to yourself in public. When every strategy gets a published postmortem — especially the ones you were excited about — cherry-picking becomes impossible. ECVT was our first real signal. We spent weeks on it. It’s dead, and we wrote that down where everyone can read it. That discipline compounds over months into something more valuable than any single strategy: the inability to self-deceive.

Other people find your bugs. Our entropy collapse strategy had a look-ahead bias in v1. In production, that’s a five-figure mistake. In public code, it’s a free review from strangers who have zero incentive to protect your ego.

And the dead teach better than the living. We learned more from FVG dying on tick data than from IPDA CE surviving walk-forward. The FVG autopsy revealed that OHLC simulation bias inflates profit factors by 4× — a finding that applies to every bar-based strategy, not just FVGs. Dead strategies generate transferable knowledge. Winners just generate P&L.

One person tests 5 strategies per month and learns from 5 failures. A community reads those postmortems and skips the same dead ends. The failure rate is 95%. Publishing that 95% is the highest-leverage thing we do.

If your edge is a mispriced option on a low-liquidity exchange — yes, shut up about it.

But our edges aren’t that. IPDA CE trades forex with structural inefficiencies that survive publication because knowing the pattern exists doesn’t help you execute it mechanically over 400 trades with proper risk management while your drawdown stretches to week three and every fiber of your brain screams “turn it off.” Cascade-Fade Scalper lives or dies on execution infrastructure — latency, fill rates, order routing — none of which travels through a GitHub link.

The hard parts of systematic trading are never the signal. They’re the boring parts: data pipelines that don’t break at 3am on a Sunday, position sizing that survives the drawdown you didn’t model, execution that doesn’t leak the edge in slippage, and the discipline to keep running a system through its losing streaks when you’re down 4% and it’s been two weeks since a winner.

Those can’t be forked. Publishing the signal is free advertising for the work that actually matters.

What Survived (Barely)

Honest accounting, March 2026:

IPDA CE — ICT dealing range with CE fill logic. PF 1.21 on 401 walk-forward trades. Live since late February, 6 trades, 50% win rate, PF 1.25. Tracking within margin of backtest. Thin edge but real — the first strategy to survive paper-to-live without blowup. Almost died once when we tested it on MT5 data with wrong parameters (min_gap_pips=0.5 instead of 3.0 — noise floor instead of signal). We nearly buried a living strategy because of a config typo. Parameter audit before funeral.

Cascade-Fade Scalper — funding rate cascade into fade entry on SOL. PF 2.56. Alive on paper, execution-dependent. BTC: dead. ETH: fragile. SOL only. Hasn’t touched a real order book yet. That PF will shrink — the question is whether it shrinks to 1.3 or to 0.8. We’ll know when we stop backtesting and start filling.

Two out of 38. And one of those might not survive deployment. This is what honest quant research looks like. Not “our proprietary algorithm generates 40% annual returns.” Two survivors, one on probation, from months of grinding.

What We Actually Publish

Full strategy specifications — entry rules, exit rules, position sizing, every parameter. Not “we use RSI and MACD with proprietary adjustments.”

Backtesting code you can run. Data sources, preprocessing, execution assumptions. If you can’t reproduce it, it’s a press release, not research.

Honest metrics — profit factor, win rate, drawdown — and the ones people conveniently omit: trade count, parameter sensitivity, out-of-sample degradation, tick-vs-OHLC comparison.

The postmortems. Every dead strategy gets an autopsy. Not “didn’t work, moving on.” A specific diagnosis — because OHLC inflation, small-sample WR bias, and walk-forward config instability are different diseases, and treating them requires knowing which one killed your patient.

What the Graveyard Taught Us

Three months of killing strategies taught us patterns no course sells, because courses sell hope and these are the opposite:

Bar-level backtests are fiction. FVG showed PF 4.28 on OHLC daily. Same strategy, same data, same parameters on ticks: PF 1.04. The only difference is honesty about when fills actually happen. If you’re not testing on tick data, you’re writing fan fiction about your returns.

Small samples are liars. Asian FVG Raid showed 70.6% win rate on 34 samples. On 234 proper UTC samples: 48.7%. Thirty-four trades told a story. Two hundred thirty-four told the truth. Never trust results on fewer than 50 trades.

Unstable configs = no edge. If your walk-forward selects different optimal parameters in every window, you’re fitting noise. The “edge” that needs monthly recalibration doesn’t exist. It never existed.

Signals don’t travel. ECVT worked on forex hourly (+198 bps) and died on equities (-88 bps). The market you test on is not the market you trade on. Always validate per-market, or you’re trading the backtest, not the market.

The hit rate is 5%. Accept this or stop doing research. Most ideas fail. The ones that survive skepticism, tick validation, walk-forward, and cross-asset testing are rare by design. Intellectual honesty means writing “dead” when it’s dead, not “needs more data” when you’re just attached to the thesis.

Thirty-eight strategies, thirty-six graves, two thin survivors, and a graveyard that teaches more than any trading course ever will.

The graveyard is open. The code is public. The corpses are documented.

That’s the product.

Start with the strategy graveyard. Code: quant-research.