polyjev
Archive · every strategy family connecting overview

Overview

Every strategy family on one line: its headline number, how many observations stand behind it, and what kind of evidence it is - live paper (real books, real prints, no real orders), a replay simulation of the journal, or a historical backtest. The status never outruns the data: below 30 observations it says exploratory; where a best variant is picked from several, its t must clear a higher bar.

  1. Loading…

Intra-candle simulator

Every journalled candle replayed against the order books recorded at each judgment. Per candle: never more than the budget out of pocket (so a candle can lose at most the budget), one side at a time, buys walk the asks and sells walk the bids, the taker fee on both, and whatever is held at the close settles at $1 or $0. Every fair-value source (the model, the stacked predictor, the model moved by the Binance lead, and a coin) is traded three ways: active (sell when the bid clears fair value - usually flipping to the other side on the same tick), flat (as active, but a flip needs a wider margin) and hold (the first buy held to settlement). Only candles the trader watched from open to cutoff are scored.

Sample

Rule parameters

Sensitivity only - choosing parameters on this same data overfits it. Changing these re-runs the simulation on the server to show how fragile a result is. The defaults are the rule under test; a parameter set picked because it looks good here has already used up this data.

Strategies

Luck band · the model against its own rule on borrowed paths

Table view

Paired comparisons

A − B per candle, over the same scored candles

Cumulative P&L by scored candle

complete, settled candles in order · cut-short candles are not in these curves · click a key to hide or show a line
Table view

Per-candle P&L

every settled candle · click a bar to open that candle
Table view

Candle

Tick table

Orders in this candle

All orders

newest first · click a row to open its candle

Paper market-maker

Sample

Variants

    Markout per share after each fill

    Table view

    Most recent fills

    Live paper trader

    The strategies in src/main.js taking paper positions on the live books, scored by src/scoreboard.js against the official outcomes. Claimed edge is what a strategy believed each share was worth at entry, minus its all-in cost; realized edge is what the share paid ($1 or $0) minus that cost. If the probability a strategy trades on is honest, the two converge over many trades - that gap, more than the P&L, decides whether to trust it.

    Strategies

    Claimed against realized edge, per share

    the horizontal gap is the number that matters

    Jev question lab

    Every question in the Jev bank is a candidate feature. src/jevlab.js admits one only if adding it to the market-anchored predictor (the book plus the exact settlement arithmetic) lowers log-loss on markets it never saw - leave one market out, each market weighted once, Holm-adjusted across the whole bank. The expected result is that none does.

    Sample

    Questions, one at a time

    change in out-of-sample log-loss per market when the question is added; positive helps

    Backtests and verification

    History the live journal cannot give yet, rebuilt from exchange data by npm run backtest, and checks that the simulators on this page find an edge that is certainly there and invent none that is not.

    The report

    Historical tests

    Historical strategies

    Equity, cumulative P&L per bet

    Verification

      Assumptions

      as the report states them