Finding one signal worth paying for.
A working page, not a product.

00The premise — is this even possible?

Yes — but only the disciplined, narrow version.

The category is proven: quant funds, paid signal groups, alt-data shops and prediction-market models all sell predictions for real money. We don't even need to beat the market for our own book — we need a track record convincing enough that people subscribe. That is a slightly lower bar than running a fund, but it is reputationally unforgiving.

★The truths most signal sellers hide

The reasons this is hard are exactly the reasons ours will be different.

◇Where we start — distil, don't search

We don't search all of finance from zero. We refine the edges our own systems already hint at.

We've attacked "find signal" three times. They're the same hunt at different resolutions — and this one is the refinery.

Three opening moves

1. Mine our own systems first. Start from signals already showing life — Monad's wallet flags have partial success, and variance in a partial win hides a real edge. Refine, don't rediscover.
2. Catalog observable cause → effect. List forced/behavioural flows we can see on-chain before price moves — exchange inflows, token unlocks, liquidation cascades, treasury/bridge moves, smart-wallet accumulation. Rank by edge × data-quality × frequency.
3. Run the portfolio through the harness. Every thesis = one screen; out-of-sample is judge; every shot logged; survivors graduate to forward-proof.

01What counts as "a signal worth paying for"

Define the target before searching, or every pretty backtest looks like a winner.

A signal isn't "it went up after X." It is a claim that clears six gates at once. We write these down first so the search has a finish line.

The scoreboard we'll rank candidates on

02The search space

The breadth is enormous. Decompose it into axes so we can narrow deliberately, not randomly.

Every candidate signal is one point across five axes. Naming the axes turns "anything in any market" into a finite map we can prune.

A · Target — what we predict

Absolute direction is the hardest and most-mined target. Relative value (A vs B, market-neutral) and volatility are structurally more predictable — we bias here.

B · Arena — where we trade

"Out of crypto" sounds like diversification but is actually harder for us: more efficient markets, less rare data. The asset-agnostic ambition is right; the realistic first hunting ground is crypto-native + prediction markets where our data is rare.

C · Source — the input that carries the edge

D · Horizon — the latency-free zone

We cannot win sub-second (no co-location, no HFT stack). We hunt hours → weeks, where the edge is in what we see and how we reason, not how fast we click.

E · Edge type — why the inefficiency exists

Behavioural (retail over/under-reaction) · Structural (forced flows, expiries, rebalances) · Informational (alt-data others ignore) · Risk-premium (paid to hold a risk). A candidate with no identifiable type is probably noise.

03Where our edge actually concentrates

The intersection of three circles — that's where we dig first.

04The anti-overfitting discipline

This isn't a section of the method. It IS the method. Break these and we ship noise.

1. Hypothesis-first, not mining-first. Start from a mechanism ("forced sellers around X cause Y"), then test it. Pure data-dredging across thousands of series guarantees false positives.
2. Train / validate / lockbox. A holdout slice is touched once, at the very end. If we peek, it's contaminated and worthless.
3. Walk-forward, not single backtest. Roll the window forward through time; a signal must keep working on data it was never fit on.
4. Cost-real from the first number. Every backtest is net of fees, slippage, market impact, and borrow. Gross edge is a fantasy.
5. Penalise the search. Track how many ideas we tried; apply a multiple-testing / deflated-Sharpe haircut. Twenty shots will produce one "winner" by luck alone.
6. Regime test. Bull, bear, chop, and at least one crisis window. An edge that only works in one regime is a beta in disguise.
7. Forward-proof before any claim. Paper-trade it live, privately, before a single public post. The backtest earns the right to a forward test; the forward test earns the right to go public.
8. The "already arbitraged?" check. If it's obvious and easy, the edge is gone. Ours must rest on data breadth or compute others don't apply.

05The funnel — from infinite ideas to one shipped signal

Eight gates. Most candidates die early and cheaply. That's the point.

06Hypothesis backlog

Concrete, mechanism-based starting bets. Each is a falsifiable claim, not a vibe.

Seed list — to be tested, killed, or promoted through the funnel. A hypothesis earns a row only if it names a mechanism, the data we already hold, and the exact target.

None of these is "the signal" — they are the first batch to run through §05. Expect most to die at Stage 4. The survivors are what we forward-proof.

A second vector — public claims to test

Thousands publicly claim signals work; almost none publish an out-of-sample test. We harvest the claim and run our gauntlet — the kills feed the graveyard, and once in a while a real one survives. Full catalog in the lab; the priority queue:

The product angle: publishing "we tested famous signal X and it's dead" is rarer and more trustworthy than another channel shouting buys.

▸Lab log — real runs, kills included

The honest record. Every screen logged, especially the ones that died.

Engine: signal-lab/ — free daily Binance data (12 assets, ~1000 days), a reusable harness with the §04 discipline baked in (time-ordered train/OOS split, cost-real, separate in-sample vs out-of-sample metrics, a persistent multiple-testing counter). A signal earns a forward-test only by clearing out-of-sample.

Screen 01 — BTC→alt-basket lead-lag (H6 family)

killed · 2026-06-14

Pre-registered: sign of BTC's k-day return predicts the alt basket outperforming BTC the next day (relative-value target, 10bps cost, 60/40 train/OOS). Tested a pre-set family k∈{1,3,7} — no cherry-picking.

k	In-sample Sharpe	OUT-OF-SAMPLE Sharpe	OOS hit-rate	OOS bps/fire
1	−0.36	−0.74	0.470	−6.92
3 headline	+0.82	−0.59	0.495	−5.46
7	+0.24	−0.43	0.500	−4.01

The lesson, live: k=3 looked like a winner in-sample (Sharpe +0.82) and bled to −24.5% out-of-sample. This is the overfitting trap from §00 caught by the engine on its very first run. Naive pure-price lead-lag is the most-mined corner of the market — exactly where we'd expect no residual edge. The hunt moves to the rarer-data hypotheses (H1 on-chain flows, H5 Polymarket) where our edge actually concentrates.

Screen 02 — Monad's invariant (mining our own system)

inconclusive · 2026-06-14

Instead of a new backtest, we read what the Monad wallet-bot already proved — its own out-of-sample validation (172k in-sample vs 74k OOS buys). Which on-chain feature actually survives OOS?

The distillation: the seductive high-win-rate features are mirages — token-age (72.7% IS) and the "gold cell" (85.2% IS) have ~zero OOS sample. The ONLY edge that replicates out-of-sample is a session-time effect: the 08–11 window beats baseline in both IS (53.7%) and OOS (54.0%, n=22,980), while 16–19 is reliably negative. Our own system already mined the wallet space — and the honest answer is no clean wallet-behaviour invariant, just a weak temporal bias (probably not enough to clear Monad's brutal costs alone). A timing filter, not a standalone signal.

Screen 03 — US-equity short-term reversal (different arena, in parallel)

killed · 2026-06-14

Run deliberately in parallel with the Monad mine, in a different arena, so we don't lock into one thing. Dollar-neutral long-losers / short-winners on 15 large caps, 5-day lookback, cost-real.

Killed: negative in-sample (Sharpe −1.6) and out-of-sample (−0.31, −9.3%). Daily reversal in large-caps is eaten by costs and dominated by momentum. A clean different-arena data point — another honest entry for the graveyard.

Screen 04 — fade funding-rate extremes (H11 · the strongest public claim)

killed · 2026-06-14

The textbook crypto claim: very positive funding = crowded longs → fade for the reversion. First pass on 66 days looked like a clear winner (OOS Sharpe +1.76). We didn't trust it. Pulling the full 3.5 years (3,150 trades) killed it: Sharpe −0.32, −39%, negative every year since 2023.

The lesson, live: the 66-day "+1.76 Sharpe" was a pure small-sample mirage — funding runs high because price is trending, so fading the trend loses. This is exactly why the harness demands data depth before any verdict. The single strongest public claim on the list, killed by the rigor the claimants skip.

07Open questions & parking lot

• What's the minimum public track-record length before charging is honest — 30 fires? 90 days? Both?
• Single flagship signal vs a small basket — basket smooths variance but dilutes the "one special thing" story.
• How do we price capacity so subscribers don't crowd out the edge they're paying for?
• Backtest data sourcing per hypothesis — what history do we have vs need to assemble?
• Manipulation defence for any social-derived signal (H4) — adversaries can spoof the input.
• Where exactly does "AI finds correlations humans miss" add value vs a human with the same data — feature discovery? regime detection? breadth of search? (This is the actual product thesis — sharpen it.)
• Longevity / decay — every edge has a half-life. How do we estimate it before going live, and build a decay-monitor that flags erosion early so we retire a signal before it embarrasses us?

⌦The graveyard — where signal goes to die

Most people only want the signals that work. Understanding why the rest don't IS the skill.

A signal dies two fundamentally different ways. Knowing which is which is the whole skill.

Death A — real signal, no edge left

Death B — there was never a signal

⌦ The headstones — signals we tested out-of-sample and buried

41 hypotheses tested · 0 deployable edges — public + academic claims, our own proprietary on-chain data, and even one signal that looked real until a deeper cut buried it (above). And our tester is calibrated: it confirmed a known-true equity anomaly, so the kills are credible — not the artifact of an always-negative test. The graveyard grows as we test more. A record of honest kills is rarer — and more trustworthy — than another channel shouting buy.