Skip to main content

Research · the evidence laboratory

Every forecastmust face the evidence.

Every model is measured against simple baselines, after costs, on data it has never seen. Across the replay, 2 of 28 strategy calls were right (95% interval 2%–23%). No edge is claimed. When nothing is good enough, the platform says so.

Recorded shadow forecast · research snapshot 27 Oct 2026 · local shadow replay, last 400 hourly bars per market. Dot: share of directional calls that were right after costs; bar: 95% interval; hollow: fewer than 30 resolved calls; dashed line: the market's always-up baseline.

03Predictive research

Each family is asked a narrow question and scored on the same history. Status comes straight from our model registry.

Gradient boosting

Will the next few bars close higher or lower?

Direction classification from engineered features — returns, volatility, volume and regime context. Boosted trees return a probability, never a price target. A scikit-learn booster stands in while the native library is integrated.

Many small trees vote; the result is a probability of “up”, not a price.Illustration — not a model output

From the model registry

  • LightGBMdirection classificationRunning in research
  • HistGradientBoosting (scikit-learn)direction classificationRunning in research

No model family has demonstrated a validated edge yet. Availability is not accuracy — “running in research” means measured, not trusted.

  • Running in research
  • Integrating
  • Needs compute
  • Licence review

04Historical market replay

Two real decisions from the shadow replay, replayed in the order they happened. What was known, what the model said, what risk allowed — and only then, what the market did.

BTC/USDT, breakout, 21 Oct 2026 · 13:00 UTC. Step 1 of 6: Market state.

BTC/USDT · 21 Oct · 13:00 UTClocked
  1. 01 · Market state

    BTC/USDT at 13:00 UTC, 21 Oct 2026

    83,826

    close of the decision hour

    Last 24 hours
    +2.04%
    Volatility (ATR 14)
    0.45% of price
    Trend (EMA 20 / 50)
    stacked up
    72h range
    80,684–83,969
  2. 02 · Information available

    Information available at that timestamp

    Only the past.

    • 72 closed hourly candlesto 13:00 UTC
    • Volume per candleBinance public spot
    • Derived featuresATR, channels, EMAs, RSI, ADX
    • News and sentimentnot connected yet
    • Social attention, macronot connected yet
    • The next 5 hoursunknowable — locked
  3. 03 · Signal generated

    Breakout model v1.0.0 · as logged

    ↑Upover the next 4 hours

    Close 83826.01 broke the prior 20-bar high 83713.91 (+0.10% buffer) with volume 2.53× average and ATR 0.44% of price.

    • Close above channel
    • Strong close
    • Volume confirmed
    • Volatility in band
  4. 04 · Independent risk decision

    Independent risk engine · it never sees the future either

    Approved

    Inside every limit — size capped by the per-position allocation.

    Sized position
    39.42 USDT
    At risk if stopped
    0.48 USDT
    Stop
    83,082 (−0.89%)
    Estimated fees
    0.08 USDT
  5. 05 · Simulated order or no trade

    Simulated order

    Paper long

    0.00047 BTC at ≈83,826

    Opened on a simulated 200 USDT paper portfolio, default limits. Paper only — no order reached an exchange. The shaded band is what the position could lose before the stop.

  6. 06 · Actual outcome, after the fact

    Keep scrolling to unlock the next 4 hours — after every decision has been made.

2 of 28

resolved calls were right across the replay window — no edge yet.

One right call and one wrong call are shown above because both happen. The scorecard is the whole window, not the highlights.

0%hit rate 7% · 95% interval 2–23%100%

local shadow replay, last 400 hourly bars per market · BTC/USDT, ETH/USDT, SOL/USDT

Opportunity discovery

The research pipeline runs every model on every bar of every market, scores each call when its horizon closes, and ranks what is left against a baseline and the cost of trading.

local shadow replay, last 400 hourly bars per market

  1. 3

    Markets scanned

    BTC · ETH · SOL

  2. 3,600

    Model evaluations

    every model, every bar

  3. 28

    Directional calls

    the rest said “no call”

  4. 28

    Resolved

    scored at the horizon

  5. 0

    Passing the bar

    ≥100 calls, above baseline, after costs

  6. Bar lengths on a log scale.

Ranked by what the evidence can support.

Each model × market, ordered by the low end of its 95% interval. To pass, a candidate needs at least 100 resolved calls, an interval wholly above the best baseline for that market, and a positive average after costs.

Baseline · BTC
21.5% always up
Baseline · ETH
33.3% momentum 24
Baseline · SOL
32.0% momentum 24
Cost hurdle
0.32% round trip

● hit rate━ 95% interval (0–100%)| baseline

  1. 01

    Breakout

    BTC/USDT · 1h

    1/3 right

    −0.34% avg net per call

    Too few calls (3 of 100 needed)

  2. 02

    Breakout

    ETH/USDT · 1h

    1/5 right

    −0.73% avg net per call

    Too few calls (5 of 100 needed)

  3. 03

    Breakout

    SOL/USDT · 1h

    0/8 right

    −0.97% avg net per call

    Too few calls (8 of 100 needed)

  4. 04

    Mean reversion

    ETH/USDT · 1h

    0/3 right

    −0.62% avg net per call

    Too few calls (3 of 100 needed)

  5. 05

    Mean reversion

    BTC/USDT · 1h

    0/2 right

    −0.35% avg net per call

    Too few calls (2 of 100 needed)

  6. 06

    Mean reversion

    SOL/USDT · 1h

    0/2 right

    −0.18% avg net per call

    Too few calls (2 of 100 needed)

  7. 07

    Trend following

    BTC/USDT · 1h

    0/2 right

    −1.22% avg net per call

    Too few calls (2 of 100 needed)

  8. 08

    Trend following

    SOL/USDT · 1h

    0/2 right

    −1.37% avg net per call

    Too few calls (2 of 100 needed)

  9. 09

    Trend following

    ETH/USDT · 1h

    0/1 right

    −1.15% avg net per call

    Too few calls (1 of 100 needed)

Avg net = mean return per resolved call after fees (0.10% per side), slippage and spread.

Nothing is ready to act on yet — and the hunter says so.

No candidate has enough calls to separate skill from luck. The scan keeps running in shadow mode; nothing graduates until the numbers do.

D · Research scoreboard

The board keepsevery result.

  1. 019Candidates
  2. 020≥ 30 resolved calls
  3. 030Interval above baseline
  4. 040Positive after costs
  5. 050Promoted
9 rows · Recorded shadow forecast
  • BreakoutSOL/USDT · 0% of 8Too few calls to judge
    95% interval
    0%–32%
    Baseline
    30%
    Coverage
    2.0%
    Net after costs
    −0.97%
  • BreakoutETH/USDT · 20% of 5Too few calls to judge
    95% interval
    4%–62%
    Baseline
    29%
    Coverage
    1.3%
    Net after costs
    −0.73%
  • BreakoutBTC/USDT · 33% of 3Too few calls to judge
    95% interval
    6%–79%
    Baseline
    21%
    Coverage
    0.8%
    Net after costs
    −0.34%
  • Mean reversionETH/USDT · 0% of 3Too few calls to judge
    95% interval
    0%–56%
    Baseline
    29%
    Coverage
    0.8%
    Net after costs
    −0.62%
  • Mean reversionBTC/USDT · 0% of 2Too few calls to judge
    95% interval
    0%–66%
    Baseline
    21%
    Coverage
    0.5%
    Net after costs
    −0.35%
  • Mean reversionSOL/USDT · 0% of 2Too few calls to judge
    95% interval
    0%–66%
    Baseline
    30%
    Coverage
    0.5%
    Net after costs
    −0.18%
  • Trend followingBTC/USDT · 0% of 2Too few calls to judge
    95% interval
    0%–66%
    Baseline
    21%
    Coverage
    0.5%
    Net after costs
    −1.22%
  • Trend followingSOL/USDT · 0% of 2Too few calls to judge
    95% interval
    0%–66%
    Baseline
    30%
    Coverage
    0.5%
    Net after costs
    −1.37%
  • Trend followingETH/USDT · 0% of 1Too few calls to judge
    95% interval
    0%–79%
    Baseline
    29%
    Coverage
    0.3%
    Net after costs
    −1.15%

Precision counts a call right only if the move after the 0.32% round-trip cost went the called way within four bars; sideways counts as wrong. The always-up baseline is the share of bars that rose by more than costs — any model has to beat it. Coverage is the share of evaluated bars on which the model made a call.

Verified performance

Each kind of evidence reported on its own: walk-forward, protected holdout and daily hash-chained shadow snapshots. No validated tradable edge.

See the evidence →

07Continuous research

Every idea runs the same loop, on data it has never seen, after costs. Most ideas fail it — and a model that fails is rejected, not tuned until it passes.

01 / 06 · one research cycle

Closed candles — and later other inputs — are stored with the time they became available, so nothing from the future can leak into the past.

Stage 1 of 6: Collect. Closed candles — and later other inputs — are stored with the time they became available, so nothing from the future can leak into the past.

How an experiment is judged.

Three gates, in order, on out-of-sample shadow results. A strategy that fails any of them is not promoted. These are the real strategies and the real scoreboard.

  • Breakout3 markets · 1h

    Resolved calls (100+ per market)

    failed:16 resolved

    Beats baseline

    failed:2 of 16 right

    Avg per call after costs

    failed:−0.78%

    Not promoted

  • Mean reversion3 markets · 1h

    Resolved calls (100+ per market)

    failed:7 resolved

    Beats baseline

    failed:0 of 7 right

    Avg per call after costs

    failed:−0.42%

    Not promoted

  • Trend following3 markets · 1h

    Resolved calls (100+ per market)

    failed:5 resolved

    Beats baseline

    failed:0 of 5 right

    Avg per call after costs

    failed:−1.27%

    Not promoted

Costs: 0.32% round trip (fees, slippage, spread). Baseline: the stronger of “always up” and 24-hour momentum in each market. local shadow replay, last 400 hourly bars per market.

Champion and challenger.

Champion
None yet
No strategy has earned it. Baselines remain the bar.
Challengers in shadow
3
Breakout · Mean reversion · Trend following
  1. AA challenger — retrained or new — is scored against the champion on the same data, costs and periods.
  2. BIt replaces the champion only if it wins out of sample. Otherwise it is rejected and logged.

Purged walk-forward validation

time →

train on the pastpurge gap · no overlapping labelstest on what came nextprotected period, scored once at the end· schematic

Retraining is not a guarantee of better results.

09Private beta

We invite researchers and traders who want to see the evidence, not a promise. You get early access and a direct line to the team; we get honest feedback.

Status
Private beta. Confirm your email and the invitation follows automatically while today's places last.
Cost
Free during the beta, with up to 50 Market Analyst questions a month.
Mode
Paper trading only. Balances, orders and results are simulated.
Access
Web terminal first. The iPhone app is in internal testing.

Your email, for one reason: to confirm your place and send your invitation. Product updates only if you opt in. No advertising trackers, and you can leave any time. Privacy notice

Request an invitation

Private beta · waitlist open

Consent

Free during the beta. Requesting an invitation creates no account and no financial relationship. Paper trading only.

You're already in.

You're signed in, so there is no need to join the waitlist.

Open dashboard

Paper trading only. Seven A.M. is a research tool — not a broker, exchange, wallet or investment advice. Forecasts can be wrong and past results do not predict future ones. Read the risk disclosure