This validation exercises the complete direct-PineScript path without copying a scraped strategy or corpus market data into this Apache-2.0 repository.

‍Historical snapshot. This run was recorded during 0.1.0 development with sampler identity pineforge_product_tpe_v1, before the finite-candidate policy work introduced the current v2 identities. The source revision and raw result JSON were not retained, so its hash is not current-release acceptance evidence. The strategy/data hashes and methodology remain useful historical validation; release claims must rerun the current executable and publish its result provenance.

Inputs

Date: 2026-07-18
Strategy: pinescript-scrapper/data/standard/shiroi-supertrend-strategy-long-short/strategy.pine
Strategy SHA-256: 1d703474287b79be82c18f90451129c912536287b32d7a9051fd53ea51a83072
OHLCV: pineforge-engine/corpus/data/ohlcv_ETH-USDT-USDT_1m.csv
OHLCV SHA-256: db8c1332da093008cfbd063e05db0b33fe8f7fd35d78cf058a366519eb9f6cc5
OHLCV range: 2020-01-01 00:00 UTC through 2026-05-04 15:14 UTC
OHLCV rows: 3,334,340 one-minute bars
Artifact key: 46cd14b22d180fb5ab8c9faf5c0ad91c272df980c935019d668571f82ee6f2fd

The Pine source was passed through pineforge_codegen.transpile_full(), compiled once, ABI/symbol checked, and reused for all trials. Tunable inputs were ATR Length as integer [3, 50] and continuous Factor [0.5, 8.0]. Enable Strategy=true and Lot Size=1.0 were fixed runtime inputs.

The study supplied input_tf=1 and script_tf=15; it did not use the derived 15-minute CSV. Every successful result reported:

input_bars_processed  = 3,334,340
script_bars_processed =   222,295
input_tf_seconds      = 60
script_tf_seconds     = 900
script_tf_ratio       = 15
needs_aggregation     = true

At the selected candidate, a one-trial control against the pre-derived 15-minute corpus file produced bit-for-bit equal objective and selected metric JSON. This confirms that the real HPO run used PineForge engine timeframe aggregation.

Study

All three samplers used seed 20260718, 4 workers, 128 trials, and:

maximize metrics.all.net_profit
       - 0.25 * metrics.equity.max_equity_drawdown
subject to metrics.all.num_trades >= 100

Best-so-far objective by equal trial budget:

Trials	dlib global	native TPE	seeded random
8	-1475.23861375	296.31731625	296.31731625
16	-324.40389625	296.31731625	296.31731625
32	-105.22328125	1618.80984625	927.27561500
64	1521.64435500	1618.80984625	955.43960125
128	1727.29872000	1787.66159250	1189.46411250

The small-budget result is intentionally retained: dlib lost to random through 32 trials on this discontinuous surface, then overtook it by 64 trials. Adaptive search is not a promise that every seed and budget beats random search.

The dlib best trial was:

trial_id            = 116
ATR Length          = 28
Factor              = 6.667564806664945
net_profit          = 2179.739380000003
max_equity_drawdown = 1809.7626399999717
num_trades          = 1335
objective           = 1727.2987200000102

Repeating all 128 dlib trials with the same seed and four-worker batch produced the exact same parameter/objective/status/metric projection. Its SHA-256 was 9e07bb65a0c04161625f2e495506d1c94c9570d1549dd39444d2dabb3f7e1233.

Native TPE validation

The same compiled strategy, 1-minute corpus, objective, constraint, seed, four-worker deterministic batch, and 128-trial budget were also run through the native TPE sampler. Its configuration was:

startup_trials = 10
ei_candidates  = 24
gamma_fraction = 0.1
gamma_cap      = 25
prior_weight   = 1.0
constant_liar  = true

All 128 trials completed with status ok and satisfied the trade-count constraint. The best trial was:

trial_id            = 105
ATR Length          = 48
Factor              = 6.8863211038926782
net_profit          = 2492.9676449999861
max_equity_drawdown = 2821.2242100000149
num_trades          = 1295
objective           = 1787.6615924999824

Every TPE trial reported the same engine aggregation evidence shown above: 3,334,340 input bars at 60 seconds became 222,295 script bars at 900 seconds, with ratio 15 and needs_aggregation=true.

Two final direct native runs took 10.00 s and 8.95 s, or 12.80 and 14.30 trials per second. Their complete result JSON files were byte-for-byte identical, including all parameters, objectives, statuses, metrics, backtest metadata, schema_version=1, pineforge_hpo_version=0.1.0, and sampler_implementation=pineforge_product_tpe_v1. Both had SHA-256 4083b7b67d001831b25aa3f4cba5c4f16dd00ebc8797869b99e41ccb4ed6f6eb.

Exploration allocation check

To test whether a strategy should always spend more of its budget on random startup exploration, the same strategy study was rerun with five fixed seeds and either 10 or 32 startup trials. Thirty-two is 25% of the 128-trial budget; all other sampler, execution, objective, constraint, and data settings were unchanged.

Seed	Startup 10 best	Startup 32 best	Winner
17	2052.878680	1869.444761	10
41	1667.607286	1072.011212	10
73	1880.846145	1659.958785	10
109	813.130771	1972.154523	32
149	1916.761325	1639.190457	10
Median	1880.846145	1659.958785	10

Startup 10 won four of five paired seeds. Every run produced 128 unique full (ATR Length, Factor) pairs, but startup 32 covered a median 41 distinct ATR values versus 35 for startup 10. Thus the larger allocation increased discrete parameter coverage without improving median objective on this two-dimensional, small-budget strategy problem.

All 1,280 trials were successful, feasible, satisfied the minimum-trade constraint, reproduced the objective from report metrics, and reported the same one-minute-to-15-minute aggregation contract. This comparison is quality evidence, not a timing benchmark. Together with the million-candidate synthetic result, it shows why exploration allocation should be compared per study rather than hard-coded globally: a rugged six-dimensional space benefited from 20% startup exploration, while this two-dimensional strategy benefited from earlier model fitting.

Wall-clock efficiency

Standalone cache-hit runs on an Apple M4 Max with 4 workers, including Python orchestration, artifact validation, parsing the 168 MB CSV once, all backtests, and JSON output, measured:

Sampler	128-trial wall time	Trials/second	Raw bar-trials/second
dlib global	9.06 s	14.13	47.11 million
native TPE	9.48 s	13.51	45.04 million
seeded random	7.21 s	17.75	59.19 million

dlib cost 25.7% and TPE cost 31.4% more wall time than random in these runs because adaptive samplers update a model and wait at deterministic four-trial batch boundaries. At 128 trials their best objectives were respectively 45.2% and 50.3% higher than this random seed. Neither percentage should be generalized without benchmarking the actual strategy, dimension count, worker count, and budget.