How OnTheRice works
OTR Signals explain world changes across General, Investments, Singapore, Social Media and Technology. Open a card for its evidence, possible consequences, uncertainty and next check. Some signals include a testable forecast. The prediction ledger reports hits divided by hits plus misses in its stated cohort; pending, unverifiable and cancelled calls are excluded under the displayed rules.
Explore Signals · Signals prediction ledger
Follow the evidence through the call
A signal separates its sourced observation, interpretation, uncertainty and next check. Read the original source and its date, then inspect the conditions that could change the view.
Investible calls keep their original direction, publication time and locked deadline. The ledger records the evidence used to check the outcome and keeps pending or unverifiable results separate from hits and misses.
Carry the lesson forward
Review outcomes without rewriting the original call. Previous ledger statistics and the current testing cohort are identified separately, and combined only in the displayed summary.
How likely is this call?
A percentage estimates the chance of a defined price move over a fixed period. It is not a promise, evidence score or past hit rate. BULLISH means expecting a rise under the stated rules; BEARISH means expecting a fall. WATCH means monitoring named evidence. A conditional call applies only if its stated trigger happens. A horizon is the forecast window: check the exact deadline. Pending means unresolved; unverifiable means the required evidence cannot be verified; cancelled means excluded for the recorded reason.
Read the public probability method
Three frameworks. One clear probability.
Researched investible calls show a percentage beside their bullish or bearish direction. If research is incomplete, the call can publish with its percentage marked research pending. Open a card for the full bullish, bearish and neutral split, its deadline, evidence and limits.
Bayesian: update from tested data where available, or state a subjective judgment based on the evidence. de Finetti: compare similar cases and account for repeated evidence. Kolmogorov: keep the three outcomes consistent and adding up to 100%.
These are uncalibrated research estimates, not proven accuracy rates. Subjective agent estimates have no verified historical sample and disclose their gaps. Earlier calls keep their original record; missing historical percentages are not filled in later.
How the percentages work → · Read the full probability reference1. Freeze the question
Each forecast names its instrument, venue, currency, reference comparison price and recorded clocks. Engine-native windows use 24, 48, 72 or 168 hours. Imported source forecasts can retain an explicitly recorded calendar-adjusted expiry: read that exact deadline and the original resolution rule. A reference price is not an executable entry price. Missing verified prices remain disclosed and must be established before scoring. A passed deadline alone does not resolve a call.
Return R = (end price − reference price) / reference price. Bullish means R > ε. Bearish means R < −ε. Neutral means −ε ≤ R ≤ ε. A flat price is not automatically bearish. This probability outcome is separate from the existing directional HIT/MISS ledger.
2. Bayesian updating and subjective estimates
Two modes are kept separate. Empirical mode computes from retained historical observations and fitted evidence where available. Subjective mode records a source-linked agent judgment, its reasoning, countercase and data gaps. It does not invent historical counts, fitted likelihoods or a Bayesian posterior. A subjective percentage is a belief estimate, not an observed market frequency.
Use historical outcomes for the same asset, horizon and regime. A Dirichlet prior combines explicit prior settings with the effective counts of bullish, bearish and neutral outcomes. Record where those counts came from.
For each state k: prior πₖ = (αₖ + effective countₖ) / Σ(α + effective counts). Then zₖ = log(πₖ) + Σⱼ wⱼ log(LRⱼₖ), and posterior pₖ = exp(zₖ) / Σ exp(z). Likelihood ratios must be fitted using data available before the forecast. When no fitted model exists, use LR = 1 and label the result a historical baseline.
Small samples can be pulled back towards the prior: η = N_eff / (N_eff + κ); final pₖ = η × posteriorₖ + (1 − η) × priorₖ. Store κ and its rationale; never hide it as a default.
3. de Finetti: fair comparisons and dependence
Exchangeability means treating selected cases as comparable under the stated conditions. It is an assumption, not a fact about every market. The empirical research record explains the asset, horizon, regime and sample selection. Repeated reports of one event form one cluster. Empirical mode stores effective sample size and dependence weights; subjective mode explains comparison limits and shared sources without claiming a historical sample.
The engine checks that an effective sample cannot exceed the raw sample. A coherent probability is not proof that a trade is profitable.
4. Kolmogorov: consistent outcomes
Every number must lie between 0% and 100%. Bullish, bearish and neutral must total 100%. Display rounding preserves that total. Probabilities at different horizons are separate forecasts; do not add them together. Normalising numbers does not make them calibrated.
5. Research before each publication
Every new investible call enters a research queue. Six roles review sources and data; historical behaviour; anomalies; narrative and behaviour; cross-asset and opposing evidence; and Bayesian synthesis. Each role records who did the work, sources, time and any data gap. Role receipts describe actual work; they do not claim six independent agents ran when they did not.
The publisher supplies actual research receipts. Empirical mode calculates from retained historical inputs; subjective mode checks an agent judgment and its source-linked rationale. The build verifies each recorded split and freezes its research and result together. Missing historical evidence, an unverified price or a missing probability does not block publication. Agents research each call automatically. Source-linked subjective estimates are allowed and clearly labelled; if no estimate is ready, the card says research pending. Integrity checks still protect recorded research from tampering and backdated evidence. Each asset on a multi-asset card needs its own percentage.
6. Check outcomes and learn
Resolve against a verified price for the same instrument, venue, currency and price basis. Keep original probabilities unchanged. Three-class Brier score = Σ(pₖ − outcomeₖ)². Log loss = −log(the probability given to the observed outcome). Evaluate later, separate data by horizon and regime, and account for overlapping windows.
This version labels all outputs “uncalibrated estimate”. It does not claim a tested calibration model, confidence interval, global accuracy level or automatic retraining. Reliability needs later forecasts and outcomes, with a held-out time period.
Preserved proposals
The PDF's example prices, percentages, likelihood ratios and thresholds remain teaching material. They are not live market inputs. Its proposed 55%, 60% and 75% display thresholds, anomaly triggers, multipliers and sample limits have not been silently adopted as trained settings. The original PDF is preserved unchanged.
Source guide
OTR Probability Intelligence Synthesis, 8 October 2026
Gneiting, Balabdaoui and Raftery (2007): Probabilistic forecasts, calibration and sharpness