Valoris — Oracle Research Terminal

VALORIS · Oracle Research

SIMULATION · SYNTHETIC DATA — model v2 · drift from 12.5k-player panel

Filter error vs TM alone · simulation

—

synthetic RMSE · real backtest: aging drift −12 % vs naive (150 players); filter not better than naive+drift

Band width (±2σ)

—

collapses on transfer anchor

Jump probability

—

particle share recently jumped

Aging drift μ(a)

—

age —

Observations

—

TM / CIES / TRANSFER

Latent value · observations · filter estimate

day 0 · window 730d · €M

latent V_t (hidden in reality) Transfermarkt (lagging) CIES (noisy, fast) realized transfer particle-filter estimate ±2σ band

speed

Jump-probability signal

filter's belief a discrete event occurred — no Kalman can emit this

Process model

Merton jump-diffusion · log state

d log V_t = μ(a_t)·dt + σ·dW_t + J_t·dN_t

aging-curve drift μ(a)deterministic lifecycle trend, estimated within-player on 212k observations: +0.89 log/yr at 18, zero at 27, −0.45 at 34. Share of upward updates falls 57 % → 30 % → 3 %.

diffusion σ·dWcontinuous form/sentiment noise around the trend. The Gaussian part a Kalman handles well.

jumps J·dNrare discrete events — injury, breakout, contract. Poisson intensity λ, multiplicative size J~N(μ_J,σ_J). The part a Kalman smears out.

Aging curve μ(a)

Δlog MV / yr · peak 27 (n = 212k)

start age 19

now: —μ(a): — log/yr

Sources · fusion weights

obs noise σ, €M

source	σ	cadence
Transfermarkt	4.0	~2–3/yr · irregular · 32 % unchanged
CIES hypothetical	6.0	snapshot only
Realized fee	1.5	sparse

static blend 52 / 28 / 20 % (methodology) · deployed contract currently encodes 50 / 30 / 20 — alignment pending · filter replaces fixed weights with likelihood weighting per observation

Editorial smoothing, measured: 32 % of Transfermarkt updates carry no value change — flat across age, league-specific (ES1 40 %, IT1 29 %). A property of the editorial desk, not the player.

Filter parameters

particles N	1 200
jump intensity λ	0.03 / day
jump size σ_J	18 €M
diffusion σ	0.6 €M/√d
resampling	systematic

What this is. A live, in-browser simulation of the Valoris oracle methodology: a hidden value process is generated, observed through three sources with their real-world failure modes (Transfermarkt's editorial smoothing lag, CIES noise, transfer-fee sparsity), and reconstructed by a particle filter. Inject an event and watch the estimate snap, or anchor a transfer and watch the confidence band collapse. Confidence band is ±2σ, matching the research convention. What it is not. Not a price feed, not real player data, not a calibrated model. Stylized: Transfermarkt is published daily here (real median cadence 144 days) and CIES appears as a continuous second source (real: one snapshot, no history). The aging drift μ(a) is the one empirical input — estimated on the 12.5k-player thesis panel. The out-of-sample test of the filter against realized fees is in progress and not yet reported. The on-chain oracle on the terminal serves the live testnet feed.