# Backtest > What Tapehouse's weekend parameters were tested against, every closure of the market since 2010, with each figure's method and the command that repeats the run. Source: https://docs.tapehouse.xyz/resources/backtest The margin engine, the weekend leverage cap, the gap backstop's limits and the gap cover's price are each set from, and then tested against, every closure of the market since 2010. This page reports one run, dated 9 October 2026, and shows how to repeat it. The run's figures are in `docs/backtest-2026-10-09.json` in the repository. ## Repeat the run The backtest reads public data and two chains, so it needs an RPC endpoint for Robinhood Chain and one for Arbitrum One. It needs Go, as set in `services/go.mod`. ```bash ROBINHOOD_RPC_URL= ARBITRUM_RPC_URL= make backtest ``` The target runs `go run ./cmd/backtest history -out backtest.json` in `services/` and writes `services/backtest.json`. Its flags are `-out`, `-deployments`, `-parameters`, `-debt-cap` and `-launch`. A second command, `backtest record `, rebuilds a deployed band from its `PriceWritten` and `Anchored` events and checks every variance, the session and every quote against the band's own views. `make test-backtest-devnode` runs that check on the local dev node. The same run feeds [calibration](https://docs.tapehouse.xyz/resources/calibration): a proposal for new parameters is refused if, over this history, it leaves more portfolios losing more than their margin, or a higher bad-debt rate, than the parameters in force. ## What was run | | | |---|---| | Assets | NVDA, TSLA, AAPL, MSFT, GOOGL, SPY | | History | 4,096 sessions, 29 June 2010 to 9 October 2026, adjusted for splits and dividends (Yahoo Finance's chart API) | | Closures | 890 weekends and holidays, from the last close to the next open; 14 of them since Robinhood Chain launched on 1 July 2026 | | Portfolios | 200 long-only and 200 long-short, 100,000 USD a unit, margined by the engine's own arithmetic at its first parameters | | Fidelity | The Go code that evaluates the band and the engine reads the Stylus programs' own test vectors, so the run uses the programs' own arithmetic | ## What the closures did The worst move of each asset from a close to the next open, in 890 closures: | Asset | Worst closure | Date | |---|---|---| | NVDA | -14.7% | 28 Jan 2019 | | TSLA | +16.8% | 17 Jan 2012 | | AAPL | -13.0% | 16 Mar 2020 | | MSFT | -11.9% | 16 Mar 2020 | | GOOGL | -10.3% | 16 Mar 2020 | | SPY | -10.4% | 16 Mar 2020 | The market is closed 31% of the time. The longest closure was 96 hours (31 October 2012, Hurricane Sandy). ## Did the margin cover them? - Across every portfolio and every closure the requirement in force covered the loss. Over the 200 long-only portfolios and the 890 closures, the requirement at the start of the ramp that rises towards a weekend covers 99.8% of the cases, and at its midpoint and its end, 100%. For the long-short portfolios the figures are 99.99%, 100% and 100%. - The ramp is needed. The weekday requirement alone, with its 25% buffer, covers the closed requirement for 193 of 200 long-only portfolios and 198 of 200 long-short ones. Against each portfolio's own worst closure it is enough for only 7 of the long-only and 178 of the long-short; the worst shortfall is 4.4%, which is what the rise towards the closed requirement before a weekend is for. - Bad debt in sixteen years: none. At the margin accounts' requirement across a closure, the engine's loss rate over the 890 closures is 0, and so is Regulation T's flat 50%. The difference is capacity: the engine lets a portfolio borrow up to 80% of its collateral's value against 50%. ## The weekend leverage cap - The cap by rule is the largest whole leverage L with 1/L at least 1.25 times the worst move of any launch asset. With a worst move of 16.8%, the rule gives 4x. - The accounts hold 5x. At 5x, 0 of the 400 portfolios lose more than their margin in any of the 890 closures, and the capacity is 80%, against 75% at 4x. The rule is the conservative figure, and the 5x deployed cap is a choice the history supports and does not prove. - Without any cap, 193 long-only and 22 long-short portfolios would have lost more than their margin, in four closures: 24 August 2015, 9 March 2020, 16 March 2020 and 5 August 2024. - For 16 March 2020, the engine counts the closure as a scenario, which leaves 88.1% capacity and costs the long-short portfolios no more margin. Handled by the cap alone, capacity is 80.0% and the long-short portfolios pay 13.2% more margin. ## Is the two-day horizon enough? The engine diffuses prices over two days. The table gives a closure's expected shortfall at 99%, squared, measured in days of a normal diffusion at each asset's volatility: | NVDA | TSLA | AAPL | MSFT | GOOGL | SPY | |---|---|---|---|---|---| | 1.85 | 1.33 | 2.70 | 1.71 | 1.38 | 2.68 | The tails of AAPL and SPY are longer than two days of diffusion. The weekend-gap scenarios and the ramp exist beside the horizon for that reason, and the coverage above, over every closure, is the test of all of them together. ## The gap backstop and the gap cover - For a debt of 100,000 USD, the exposure limits the history asks for are 954 USD for the cross positions and 4,079, 4,260, 2,079 and 838 USD for NVDA, TSLA, AAPL and MSFT; GOOGL and SPY need none. In 5.07 years of closed market the backstop would have absorbed 1,635 USD, so a weekend premium of 36 basis points a year pays for it. See [the gap backstop](https://docs.tapehouse.xyz/contracts/gap-backstop). - The gap cover's price was checked on weekends it was not fitted on. In 61% of the 890 weekends the gap the cover prices was set by the prior week's move, with the floor not binding. The 24/5 reopening's realised move exceeded the priced gap in 0.17% of asset-weekends under uniform demand, 0.06% when demand follows the prior week's move, and 0.02% when it follows its square; the engine's gap alone would have been exceeded 0.09% of the time. On the 14 weekends since launch, never. See [weekend gap cover](https://docs.tapehouse.xyz/concepts/weekend-gap-cover). ## The band against the reopening For each closure since launch, the run recomputes the band as it would have quoted when the closure's last public 24/7 value was written, and checks whether the first Chainlink round after the reopening fell inside it. RedStone keeps public 24/7 values for 30 days, so only 24 of the 84 asset-closures have one; the rest are counted only. | Chain | Reopenings evaluated | Inside the band | Median error | Median lead | |---|---|---|---|---| | Robinhood Chain | 24 of 84 | 19 | 15.4 bps | 3,580 s | | Arbitrum One | 24 of 84 | 11 | 40.2 bps | 1,796 s | That is 79% on Robinhood Chain and 46% on Arbitrum One, on small samples. The band is the price with a declared doubt, written minutes before the close; what lies beyond it is what the margin, the backstop and the [reopening auction](https://docs.tapehouse.xyz/contracts/reopening-auction) are for. From deployment on, the band's own `PriceWritten` events keep every closure, so the sample grows with each weekend. See [the band](https://docs.tapehouse.xyz/concepts/band). ## What it costs to liquidate The liquidator's price falls 60 basis points a minute with the market open, reaching its floor in 16.7 minutes, and 15 a minute with it closed, in 66.7. The next table compares, at block 87,026,892 on Robinhood Chain, the engine's linear liquidity charge with what each pool's quoter would have cost, at a tenth, half and all of the depth: | Asset | Engine | Pool | |---|---|---| | NVDA | 0.55% / 2.55% / 5.05% | 0.96% / 32.51% / 66.26% | | TSLA | 0.80% / 2.80% / 5.30% | 0.50% / 1.57% / 2.94% | | AAPL | 0.55% / 2.55% / 5.05% | 0.16% / 0.86% / 1.90% | | MSFT | 0.80% / 2.80% / 5.30% | 1.05% / 3.59% / 6.53% | | GOOGL | 0.55% / 2.55% / 5.05% | 0.11% / 0.99% / 2.06% | | SPY | 0.55% / 2.55% / 5.05% | 0.28% / 0.97% / 1.74% | The charge covers the pools of TSLA, AAPL, GOOGL and SPY. It does not cover NVDA's, which cost far more, nor MSFT's at any of the three depths, by 0.25 to 1.2 points, at that block. The calibrator re-measures that figure, which is why a depth falls to its measured value at once while it rises at most 1.5 times a step. Robinhood Chain's mainnet pools are not ours, and the testnet's are small. ## What this does not show - A backtest is history. A closure worse than every one since 2010 would still be a bad debt; the backstop and the auction exist for it. - The portfolios are synthetic: two hundred of each kind, at 100,000 USD a unit, with no real users among them. - The run covers six assets, the launch set. The testnet runs two of them. - The band's accuracy is measured on 24 asset-closures per chain, a far smaller sample than the decade behind the other figures. ## Next - [Calibration](https://docs.tapehouse.xyz/resources/calibration) - [Liquidation](https://docs.tapehouse.xyz/concepts/liquidation) - [Portfolio margin](https://docs.tapehouse.xyz/concepts/portfolio-margin)