{
 "generated": "2026-08-03 21:32 UTC",
 "shipped_config": "T0.3 s1",
 "config": "Kronos-mini, tokenizer-2k, context 256, T0.3 s1",
 "method": "Walk-forward backtest on daily candles with NON-OVERLAPPING windows (step = horizon), up to 40 windows per asset, and 110-120 windows per asset for the four classes measured deeply (crypto, equity, forex, commodities), horizon 7 days (crypto/equity/forex) and 5 days (commodities). Direction = sign of the predicted move vs realised. mae_pct is the average absolute error of the predicted move; rw_mae_pct is the same for a random walk (predicting no change at all) \u2014 if the model does not beat that, its point forecast adds nothing. Windows before 2025-08-02 overlap Kronos's own pretraining data, so post_release is the conservative read.",
 "why_per_class": "Published per asset class because one pooled number, repeated on four pages, said nothing about which market the model can or cannot read.",
 "categories": {
  "commodities": {
   "n": 1200,
   "assets": 10,
   "direction_hit_rate_pct": 49.2,
   "ci95": [
    46.4,
    52.1
   ],
   "mae_pct": 5.038,
   "rw_mae_pct": 4.179,
   "beats_random_walk": false,
   "ic": 0.02,
   "net_edge_per_trade_pct": -0.2074,
   "t_clustered": -1.18,
   "trading_days": 534,
   "per_asset_hit_pct": {
    "CL=F": 47.5,
    "GC=F": 54.2,
    "HG=F": 54.2,
    "KC=F": 40.8,
    "NG=F": 48.3,
    "PA=F": 50.0,
    "PL=F": 53.3,
    "SI=F": 53.3,
    "ZC=F": 48.3,
    "ZW=F": 42.5
   },
   "deep_study": {
    "question": "Does Kronos have a measurable directional edge on commodities?",
    "pre_committed_bar": "direction >= 52% AND 95% CI lower bound > 50",
    "windows_per_asset": 120,
    "assets": 10,
    "passes_bar": false,
    "note": "Re-measured on 120 non-overlapping windows per asset (1200 predictions) instead of 40 (280). Equity and crypto had already been measured this deeply; these two had not, and at 280 predictions the confidence range was +/-5.5 points \u2014 too wide to call either way. It is now +/-2.8 points.",
    "previous_estimate_pct": 51.1,
    "previous_n": 280
   }
  },
  "crypto": {
   "n": 880,
   "direction_hit_rate_pct": 53.4,
   "ci95": [
    50.1,
    56.7
   ],
   "mae_pct": 7.28,
   "rw_mae_pct": 6.73,
   "beats_random_walk": false,
   "ic": 0.035,
   "assets": 8,
   "sample_note": "Re-measured on 110 non-overlapping windows per asset (880 predictions) instead of 40 (320). The earlier figure of 55.9% was not wrong on its own sample \u2014 the same 40-window slice still reads 55.6% \u2014 but it was the most recent slice, and the wider sample is the better estimate.",
   "tradeable": false,
   "tradeability_note": "Measured directly: traded with market entry (no limit, so no adverse selection), volatility sizing and 20 bps round-trip costs, this edge nets +0.003%% per trade \u2014 Sharpe 0.005, t=0.03 over 880 trades. A 53%% hit rate at a 7-day horizon with symmetric payoffs does not survive costs. Real, but not tradeable.",
   "per_asset_hit_pct": {
    "ADA/USDT": 57.3,
    "AVAX/USDT": 56.4,
    "BTC/USDT": 53.6,
    "DOGE/USDT": 46.4,
    "ETH/USDT": 55.5,
    "LTC/USDT": 53.6,
    "SOL/USDT": 50.9,
    "XRP/USDT": 53.6
   }
  },
  "equity": {
   "n": 1800,
   "direction_hit_rate_pct": 51.5,
   "ci95": [
    47.8,
    55.0
   ],
   "mae_pct": 4.18,
   "rw_mae_pct": 3.76,
   "beats_random_walk": false,
   "ic": 0.015,
   "assets": 8,
   "deep_study": {
    "question": "Does Kronos have a measurable directional edge on US equities?",
    "pre_committed_bar": "direction >= 52% AND 95% CI lower bound > 50",
    "configs": {
     "ctx256 T0.3 s1": {
      "n": 900,
      "direction_hit_rate_pct": 51.1,
      "ci95": [
       47.8,
       54.4
      ],
      "mae_pct": 4.15,
      "rw_mae_pct": 3.77,
      "ic": -0.041,
      "passes_bar": false
     },
     "ctx512 T0.3 s1": {
      "n": 900,
      "direction_hit_rate_pct": 51.8,
      "ci95": [
       48.5,
       55.0
      ],
      "mae_pct": 4.22,
      "rw_mae_pct": 3.75,
      "ic": 0.004,
      "passes_bar": false
     }
    },
    "verdict": "No configuration cleared the bar: Kronos shows no measurable directional edge on equities. The site should say so rather than show a number that looks like skill.",
    "assets": 15,
    "windows_per_asset": 60
   },
   "conclusion": "No measurable directional edge on equities. Tested again with a pre-committed bar (>=52% and a 95% lower bound above 50) on 1,800 windows across 15 large caps, at two context lengths: 51.1% and 51.8%, both confidence ranges straddling 50, and the point forecast still worse than assuming no change. Treat the Kronos column on equities as decoration until this changes."
  },
  "forex": {
   "n": 1200,
   "assets": 10,
   "direction_hit_rate_pct": 51.7,
   "ci95": [
    48.8,
    54.5
   ],
   "mae_pct": 1.075,
   "rw_mae_pct": 0.969,
   "beats_random_walk": false,
   "ic": 0.022,
   "net_edge_per_trade_pct": 0.0594,
   "t_clustered": 0.32,
   "trading_days": 360,
   "per_asset_hit_pct": {
    "AUDJPY=X": 49.2,
    "AUDUSD=X": 43.3,
    "EURGBP=X": 60.0,
    "EURJPY=X": 43.3,
    "EURUSD=X": 50.8,
    "GBPUSD=X": 64.2,
    "NZDUSD=X": 49.2,
    "USDCAD=X": 55.8,
    "USDCHF=X": 52.5,
    "USDJPY=X": 48.3
   },
   "deep_study": {
    "question": "Does Kronos have a measurable directional edge on forex?",
    "pre_committed_bar": "direction >= 52% AND 95% CI lower bound > 50",
    "windows_per_asset": 120,
    "assets": 10,
    "passes_bar": false,
    "note": "Re-measured on 120 non-overlapping windows per asset (1200 predictions) instead of 40 (320). Equity and crypto had already been measured this deeply; these two had not, and at 320 predictions the confidence range was +/-5.5 points \u2014 too wide to call either way. It is now +/-2.8 points.",
    "previous_estimate_pct": 51.9,
    "previous_n": 320
   }
  },
  "overall": {
   "n": 1240,
   "direction_hit_rate_pct": 52.6,
   "ci95": [
    49.8,
    55.3
   ],
   "mae_pct": 4.63,
   "rw_mae_pct": 4.2,
   "beats_random_walk": false,
   "ic": 0.023,
   "assets": 31
  },
  "post_release": {
   "n": 1160,
   "direction_hit_rate_pct": 52.9,
   "ci95": [
    50.1,
    55.8
   ],
   "mae_pct": 4.75,
   "rw_mae_pct": 4.3,
   "beats_random_walk": false,
   "ic": 0.023,
   "assets": 31
  }
 },
 "configs_measured": {
  "T0.3 s1": {
   "n": 1240,
   "direction_hit_rate_pct": 52.6,
   "ci95": [
    49.8,
    55.3
   ],
   "mae_pct": 4.63,
   "rw_mae_pct": 4.2,
   "beats_random_walk": false,
   "ic": 0.023,
   "assets": 31
  },
  "T0.6 s1": {
   "n": 1240,
   "direction_hit_rate_pct": 52.0,
   "ci95": [
    49.2,
    54.8
   ],
   "mae_pct": 5.01,
   "rw_mae_pct": 4.2,
   "beats_random_walk": false,
   "ic": -0.001,
   "assets": 31
  },
  "T0.6 s5": {
   "n": 1240,
   "direction_hit_rate_pct": 53.1,
   "ci95": [
    50.3,
    55.8
   ],
   "mae_pct": 4.59,
   "rw_mae_pct": 4.2,
   "beats_random_walk": false,
   "ic": -0.011,
   "assets": 31
  },
  "T1.0 s5": {
   "n": 1240,
   "direction_hit_rate_pct": 52.2,
   "ci95": [
    49.4,
    54.9
   ],
   "mae_pct": 4.76,
   "rw_mae_pct": 4.2,
   "beats_random_walk": false,
   "ic": 0.017,
   "assets": 31
  },
  "prod T1.0 s1": {
   "n": 1240,
   "direction_hit_rate_pct": 50.5,
   "ci95": [
    47.7,
    53.3
   ],
   "mae_pct": 5.58,
   "rw_mae_pct": 4.2,
   "beats_random_walk": false,
   "ic": 0.027,
   "assets": 31
  }
 }
}