N-day script timeframe built from finer input bars: last_bar_index counts input days, the open trade is missing from the report, and 3D bars follow a different grid than TradingView's #811

Summary

When the input bars are finer than an N-day script timeframe (input D or 240, script 2D, 3D or 5D), the engine builds the chart bars itself. Two things then differ from passing N-day bars directly:

1. last_bar_index is 186 on every bar, while the last chart bar's bar_index is 61 (3D), 93 (2D) or 37 (5D); the run has 188 input days. barstate.islast is never true. The position still open at the end has no row in the trade list: net_profit is the sum of closed trades, while the last equity point includes the open position.
2. The N-day bars fall on a grid counted from 1970-01-01. Around the new year the 3D bars open on 2023-12-25, 12-28, 12-31, 2024-01-03 and 01-06, so one bar spans the year boundary. The engine's own request.security(syminfo.tickerid, "3D", time) on a daily chart restarts on 1 January (2023-12-24, 12-27, 12-30, 2024-01-01, 01-04, 01-07; the 12-30 bar is two days long), which src/source/multi_day_periods.hpp describes as TradingView's rule, and TradingView's 3D chart of BINANCE:BTCUSDT does the same. So the same script gives different trades depending on whether the host passes 3D bars or finer bars. 2D and 5D show the same 1970 grid (5D: 2023-12-24, 12-29, 2024-01-03).

Input D with script W, monthly from daily, and input = script = 3D behave as expected.

Versions

pineforge-engine v1.11.0 (a4472324) with pineforge-codegen 1.11.0, and v1.10.0 (59f98ba8) with Codegen 1.10.0: identical results on both. Apple clang 17.0.0, macOS 15.8 arm64. BTCUSDT daily bars from 2023-12-01 to 2024-06-05 (188 bars; the 4h run uses 1,123 4h bars over the same dates; the 3D / 3D run uses 63 3D bars of the same data that restart on 1 January, opening 2023-12-03 to 2024-06-05, the first bucket with only 12-01 and 12-02 left out), initial capital 10000, no commission or slippage, run_backtest_full_checked with bar_magnifier = 0. Values come from // @pf-trace and the run report.

TradingView: BINANCE:BTCUSDT, 3D, Deep Backtesting from 2023-12-01 to 2024-06-05, measured on 2026-11-10.

Reproduction

//@version=6
strategy("3D chart from daily bars", overlay = true, initial_capital = 10000, default_qty_type = strategy.percent_of_equity, default_qty_value = 10)
fast = ta.ema(close, 3)
slow = ta.ema(close, 8)
if ta.crossover(fast, slow)
    strategy.entry("Long", strategy.long)
if ta.crossunder(fast, slow)
    strategy.entry("Short", strategy.short)
last_index = last_bar_index
is_last = barstate.islast ? 1 : 0
bar_open = time
// @pf-trace last_index=last_index
// @pf-trace is_last=is_last
// @pf-trace bar_open=bar_open

Entries (date, side):

Each TradingView entry date is day 1 + 3k of 2024. In the D / 3D run, the Long opened on 2024-05-17 is still open on the last chart bar but has no row in the trade list; it counts only in equity (10387.63 − 10331.05 = 56.58).

The request.security grid, with input = script = D:

//@version=6
strategy("3D bars requested from a daily chart", overlay = true, initial_capital = 10000)
requested_open = request.security(syminfo.tickerid, "3D", time)
// @pf-trace requested_open=requested_open

From a read of the source, the count of expected script bars in PineScheduler::run_begin (src/source/pine_scheduler_native.cpp) comes out near the number of input days for these timeframes, so the run never reaches the bar it treats as last. The 2D run's last chart bar (2024-06-04, covering 06-04 and 06-05) is complete, so this is not the unfinished last bucket that is skipped on purpose.

Expected

last_bar_index equals the last chart bar's bar_index, barstate.islast is true on it, and the open position is in the trade list. N-day chart bars built from finer input restart on 1 January, like the engine's request.security N-day bars and TradingView's chart, so passing daily bars or 3D bars gives the same trades.

Workaround

Aggregate the input to N-day bars in the host, restarting on 1 January, and pass input = script timeframe. With the bar magnifier this loses the finer path.