AI Decision
SimulatedKronos Forecast
SimulatedAccount & Risk
Open Positions
Simulation ledgerForecast Performance
Control Center
Pause blocks new entries. Protective monitoring continues. Close requests flattening and pauses entries.
Event Log
Local timeForecast Performance
Forecast quality and trading outcomes are separate measurements.
Review a recorded forecast
Each forecast is permanently recorded before its outcome is scored. Select one to compare its paths with subsequently observed candles.
Metric definitions
Directional accuracy: terminal mean forecast and observed terminal close move in the same direction from the input close. Zero-direction outcomes count as incorrect. Price error: mean absolute forecast-close error over each finished horizon, in pips. Only complete observed targets qualify.
Trading win rate: closed trades with positive net P/L divided by closed trades. Net result: realised simulation P/L in USD; includes bid/ask spread. Replay includes separately configured slippage. All figures include their sample size and horizon. No arbitrary composite forecast score is used.
Trades & Decisions
Persistent simulation records. OANDA Practice and real-money execution are unavailable.
Trades
Order intents
Decisions & executions
Historical Replay
Chronological forecasts, revealed outcomes and conservative fill assumptions.
Dataset library
Import completed M1 candles with time,open,high,low,close. Time must be UTC epoch seconds or timezone-qualified ISO 8601. Import creates a separate dataset; it never replaces the live ledger.
Start a bounded replay
Consensus requires a reversal of the previous qualifying terminal-close direction. Both modes use identical data and costs, 0.25% stop risk and a 2% daily equity gate. Ambiguous stop/target candles use the stop first. Targets crossing split boundaries are excluded. Final test data never selects or promotes a model.
Training Lab / Model Versions
Scoring is functional. Weight updates are simulated and clearly labelled.
Bounded training-interface demonstration
This job records chronological forecasts, finished target scores and simulated checkpoints. It does not train Kronos, produce model weights, claim improvements, or promote a model.
Model registry
Pretrained data coverage is unknown. Deployed and candidate records are separate. Real model promotion and Practice evaluation require later approval; no activation action is exposed in this build.
Connections & Settings
One authoritative configuration, enforced again by the execution gateway.
Provider connections
Session, analysis & risk
Proposed 0.25% risk means equity at the protective stop, not position value or margin. Practice trading requires your later confirmation of risk, daily shutdown and closure policy.
Daily baseline: current equity at local midnight in the selected timezone. Default end policy: block entries, cancel pending plans, request closure and keep the watchdog running. Ordinary stops and loss thresholds do not guarantee a maximum loss.
System Health / AI Usage
Measured latency and provider usage from persistent records.
Stage latency
Simulation fault controls
Inject a dependency failure to exercise entry blocking and protective monitoring. Clearing faults does not retry an uncertain order.
Connection boundaries
The browser displays server state. The worker continues after this tab closes. Simulation protective stops are gateway-held; broker-held stops require actual OANDA qualification. External provider credentials are never part of AI requests.