Past data only
Every validation forecast uses data up to its forecast time. Model parameters are refitted as the window moves.
Validation
A model enters Labs only after an out-of-sample check on real ArcDelta data. The validation figures stay next to every result.
Compared with the outcome
The forecast is compared with what happened next: did the price land in the band, was the loss tail breached, was the level touched.
Metrics
50% and 90% band coverage, VaR breach rate with the Kupiec test, Brier score for touch probabilities, median error for costs, false-positive rate for backtest statistics.
Live calibration
Inside Monte Carlo, VaR and Liquidation risk the same test is repeated on the selected instrument over the last 10 days.
Periodic review
Every two weeks results are recomputed on new data and the verdict is published on the model card in Research.
Market modelsv2.0 · Beta
Liquidation map
At which prices are leveraged longs and shorts likely to be liquidated?
Held-out days 13–19 October, 30 perpetual contracts. Price crossing version 2.0 levels explains open-interest closes 3.4 times better than the previous version, with a positive effect on all 30 instruments. About 9% of a level's estimate closes when price crosses it (90% interval 4–13%). Reviewed every 14 days.
Market modelsv2.0 · Validated
Monte Carlo
What range will the price cover, and how likely is it to touch a given level?
Horizons longer than one day run engine 2.0: a daily volatility level (HAR with the last 4h and 1h), 5m steps for the first day, then 1h and 4h steps. Out of sample on 30 perpetuals, January 2025 to October 2026, with the 27 days of history ArcDelta holds: CRPS 0.9% better than 1.1 at 7 days and 0.4% at 3 days, the 7-day upper 1% tail breached 1.4% of the time (1.1: 1.9%), and the run is about four times faster. Up to one day and for indicator distributions engine 1.1 runs. Reviewed every 14 days.
Risk modelsv1.1 · Beta
Value at Risk
How much can this position lose over the next hour, day or week?
Same engine as Monte Carlo 1.1. Out of sample on 30 contracts from 1 September to 18 October 2026 the one-hour 5% loss tail was breached 4.4% of the time and the 1% tail 1.1%; at four hours 5.1% and 1.1%.
Risk modelsv1.1 · Beta
Liquidation risk
How likely is my leveraged position to be liquidated, and what leverage keeps that risk low?
Touch probabilities include intrabar highs and lows. On 30 perpetuals the predicted probability differed from the observed touch frequency by 1.8 points on average at 1h, 1.0 at 4h and 2.5 at 24h (engine 1.0 check; 1.1 bounds the variance reaction to single outlier bars).
Risk modelsv1.0 · Beta
Trade plan
How likely is my take to fill before my stop, and what is the trade worth in R?
On 28 perpetuals, about 27,000 setups per horizon, the predicted take-first probability differed from the observed frequency by 2.4 points on average at 1h, 1.3 at 4h and 3.3 at 24h.
Market modelsv1.0 · Beta
Volatility forecast
How much will this instrument move over the next hour, four hours and day, and are options pricing more or less than that?
On a held-out week across 28 perpetuals, HAR had a QLIKE loss 4.7% lower than GARCH at 1h, 3.3% at 4h and 3.7% at 24h; it won on 20, 17 and 13 of the 28 instruments.
Market modelsv1.0 · Beta
Market structure
How tied is each coin to BTC, in calm and in stress, and which coins move together once BTC is taken out?
On 134 perpetuals over 8 days the 134 × 134 matrix took under 0.2 s for every window. Between consecutive days the ranking of coins by correlation to BTC held (Spearman 0.77), while clusters agreed only partly (adjusted Rand 0.26).
Market modelsv1.0 · Beta
Seasonality
Which hours of the day and days of the week are really different for this instrument?
Over 20 days on 28 perpetuals, 376 of 3,276 tests passed the false-discovery control at q = 0.10: higher turnover and volatility at the US open (13:00–15:00 UTC) on 26 and 25 instruments, quieter weekends on about half. Return effects passed 6 times in total and the up-hour share never.
Market modelsv1.0 · Beta
Hawkes order flow
How much of the current trading is caused by earlier trading, and is the flow heating up?
On simulated flow with known parameters the fit recovers the branching ratio within 0.15, the residuals pass the KS test and a plain Poisson fails it. A live check on real trades appears on this page for every run.
Market modelsv0.1 · Beta
Market-maker sandbox
What would a simple two-sided quoting strategy earn or lose over the next minutes if the flow keeps its Hawkes clustering, calm and stressed?
Simulation only. The PnL split is exact (the three parts sum to the total) and is covered by tests; the price model is not validated against real order books.
Market modelsv1.0 · Beta
Funding carry
What will holding this perpetual cost or pay in funding over the next day, on each venue?
0.6 × last settled rate + 0.4 × 3-day mean predicted the next-day average rate with an MAE of 0.51 bp, against 0.59 bp for the last rate and 0.78 bp for zero, on 488 venue symbols.
Backtest modelsv1.1 · Beta
Execution cost
What will a market order of my size really cost on each venue?
The trailing two-hour median predicted the cost one hour later with a median error of 6.6% ($10k) to 9.3% ($500k). Version 1.1 band: on the held-out day, 23 October, 79.8% of future costs fell inside the p10–p90 band against an 80% target (the old band 65.7%); BTC and ETH are still under-covered.
Backtest modelsv1.0 · Validated
Backtest check
Is this backtest Sharpe real, or the best of many lucky tries?
In 1,000 null simulations the best of 50 variants passed plain PSR 92% of the time and DSR 0%; a real edge among 20 variants was found 37% of the time.