{
 "generated": "2026-08-03",
 "what": "The model's own directional output (the `lean` it archives every run, and the number the BUY/SELL sections are derived from) scored against the price that actually happened, on every archived snapshot whose 7-day window has closed. Not a simulation of the model: its real output, replayed.",
 "method": "One call per asset per generation day (intraday regenerations are the same view, not new evidence). Direction is measured from the published spot to the close at the horizon. Expectancy is net of round-trip costs by asset class.",
 "overall": {
  "n": 123,
  "assets": 62,
  "period": [
   "2026-07-23",
   "2026-07-27"
  ],
  "direction_hit_rate_pct": 59.3,
  "ci95": [
   50.5,
   67.6
  ],
  "net_expectancy_pct": 0.828,
  "t_stat": 1.75
 },
 "by_class": {
  "commodities": {
   "n": 7,
   "hit_pct": 28.6,
   "ci95": [
    8.2,
    64.1
   ],
   "net_expectancy_pct": -1.793
  },
  "crypto": {
   "n": 9,
   "hit_pct": 88.9,
   "ci95": [
    56.5,
    98.0
   ],
   "net_expectancy_pct": 3.453
  },
  "equity": {
   "n": 59,
   "hit_pct": 59.3,
   "ci95": [
    46.6,
    70.9
   ],
   "net_expectancy_pct": 1.14
  },
  "forex": {
   "n": 48,
   "hit_pct": 58.3,
   "ci95": [
    44.3,
    71.2
   ],
   "net_expectancy_pct": 0.335
  }
 },
 "reading": "Two samples, two answers. Our archived calls scored 59.3% over 123 windows in a single week. A replica of the same model family, run over a decade on 7,823 non-overlapping windows, scored 50.2% (95% range 49.1-51.3) with an information coefficient of -0.006. The large sample is 64 times the small one and finds nothing. The honest reading is that the 59.3% is most likely a one-week artefact, and it stays on this page only because the live ledger \u2014 not our preference \u2014 gets to settle it.",
 "caveats": [
  "A decade-long replica of the same model family found NO edge (50.2%, n=7,823). Read the 59.3% against that, not on its own.",
  "123 windows across one week (23-27 July 2026) \u2014 one market regime, not a cycle.",
  "Crypto's 88.9% is 9 observations and should be read as noise, not skill.",
  "Commodities scored 28.6% on 7 observations \u2014 equally uninformative, and shown for symmetry.",
  "Windows from consecutive days overlap, so the effective sample is smaller than 123.",
  "The window covers 23-27 July; the BUY/SELL presentation went live on 27 July, so the earlier calls were archived and scoreable but not yet displayed as such."
 ],
 "family_backtest": {
  "what": "A model of the SAME family \u2014 same feature families (momentum, volatility, trend), same 7-day target, walk-forward with the overlapping horizon purged from training \u2014 run over a decade on 27 assets. It is a replica, not our production pipeline (which also uses macro, fundamental and cross-asset inputs), so it cannot prove our model has no edge. It can show whether this class of model has one.",
  "non_overlapping": {
   "label": "non-overlapping (step = horizon)",
   "n": 7823,
   "direction_hit_rate_pct": 50.2,
   "ci95": [
    49.1,
    51.3
   ],
   "ic": -0.0062,
   "gross_expectancy_pct": 0.037,
   "net_expectancy_pct": -0.057,
   "t_stat": -0.63
  },
  "by_class": {
   "commodities": {
    "n": 1904,
    "direction_hit_rate_pct": 50.3,
    "ci95": [
     48.1,
     52.6
    ],
    "net_expectancy_pct": 0.053
   },
   "crypto": {
    "n": 2040,
    "direction_hit_rate_pct": 48.6,
    "ci95": [
     46.5,
     50.8
    ],
    "net_expectancy_pct": -0.396
   },
   "equity": {
    "n": 2175,
    "direction_hit_rate_pct": 51.1,
    "ci95": [
     49.0,
     53.2
    ],
    "net_expectancy_pct": 0.117
   },
   "forex": {
    "n": 1704,
    "direction_hit_rate_pct": 50.8,
    "ci95": [
     48.4,
     53.1
    ],
    "net_expectancy_pct": 0.003
   }
  },
  "verdict": "No directional edge. 50.2% over 7,823 non-overlapping windows, IC -0.006, t -0.63."
 }
}