tradeblocks-mcp 4.2.0 → 4.4.0

package/README.md

@@ -537,9 +537,80 @@ predictive reports, regime analysis and threshold sweeps. `import_daily_dataset`
 atomically, without changing market data or requiring provider credentials. Date-only close
 values are available the following calendar day at 00:00 ET; date-only open values require
 `knownBy`; timestamped rows are as-of. Sparse columns carry forward their latest available
-observation without a freshness cutoff. See the [dataset tools reference](../../docs/mcp-tools.md#market-analysis)
+observation with no freshness cutoff by default. See the [dataset tools reference](../../docs/mcp-tools.md#market-analysis)
 and [4.1.0 behavior changes](../../releases/v4.1.0.md#behavior-changes).
 
+With Docker's default `-v ./data:/data` mount, put the CSV at
+`./data/datasets/vrp.csv` on the host and pass `csvPath: "datasets/vrp.csv"` or
+`"/data/datasets/vrp.csv"`; use the container path beneath the configured data root.
+Close-column metadata shows `knownBy: null`; open columns retain the configured value.
+Field-based `suggest_filters` retains all-trade `currentStats` and reports per-field
+`fieldBaselines` (`count`, `totalPl`, `uncoveredCount`, `uncoveredTotalPl`).
+Suggestions compare against those covered populations; `filter_curve.baseline`
+reports the same covered and uncovered counts and P&L.
+
+New in 4.4.0: [filter statistics](../../releases/v4.4.0.md) for asking whether a
+chosen filter survives a full threshold search, with explicit assumptions and work caps.
+
+`find_predictive_fields.method` selects `pearson` (default), `spearman`, or `kendall`
+(tau-b with ties). Explicit requests name their method; omission preserves the Pearson
+response. Ranking uses the chosen coefficient on the same causal entry-time values,
+including `maxMissedSessions` refusals.
+Kendall admits up to 75 million estimated pairs across fields, using each field's
+valid field/target coverage after strategy/date filtering. It refuses by name
+(`kendall_pair_work_limit`) before quadratic calculations. A 1,100-trade block with
+100 fully covered fields passes (60,445,000 pairs); a 20,000-trade field is refused.
+Pearson and Spearman retain their existing behavior.
+
+`filter_curve` accepts either the existing `field` or up to 10 distinct `fields`.
+Multi-field calls, an explicit `objective` (`totalPl` or `winRate`), or opt-in
+`permutation: { iterations: 499, block: "month", seed: 42 }` add the best threshold
+and per-year kept/removed stability. Permutation shuffles whole calendar-month outcome
+blocks and reruns the complete threshold sweep, not just the selected threshold.
+It reports add-one raw p-values and Holm-adjusted p-values across requested fields.
+Single-field calls without these options retain their old outputs.
+At least 30 covered trades and 8 observed months are required; refusals are named.
+Iterations are capped at 499. Only permutation requests limit threshold and percentile
+arrays to 100 entries, with a named `permutation_candidate_limit` refusal.
+Without permutation, larger arrays retain their previous schema acceptance.
+Sort-aware work admission caps permutation at 30 million units per field and
+90 million per request, refusing the whole request before any sweeps. At 1,100 trades,
+55 months, 499 iterations and default thresholds, the largest admitted family is
+3 fully covered fields (87,751,152 estimated units); 4 fields exceed the cap.
+Field, objective and mode choices across separate requests are outside the Holm family.
+Monthly blocks retain within-month dependence, not cross-month dependence; block
+exchangeability is an assumption. See [the full contract](../../docs/mcp-tools.md#filter-curve-significance-and-stability).
+
+`suggest_filters` defaults to `objective: "totalPl"`: only positive P&L improvements,
+ranked by P&L improvement with win-rate gain as tie-breaker. `objective: "winRate"`
+admits/ranks by win-rate gain with P&L tie-breaks. `minImprovementPct` (default 3
+percentage points) applies only to `winRate`; it is ignored under `totalPl` and the
+response notes this when supplied. Standalone and composite suggestions include
+`tradesKept`, `keptFraction`, `tradesRemoved`, and `objective`. `confidence` means
+sample-size support, not statistical confidence; keeping less than 20% caps it at
+low and adds a low-retention warning. Composite metrics describe kept trades.
+
+In 4.3.0, `minImprovementThreshold` is null under `totalPl` and numeric under
+`winRate`. Shared support thresholds are high at 20 kept/10 removed, medium at
+10 kept/5 removed, otherwise low, subject to the retention cap. This replaces
+the fields path's former kept-only 100/30 thresholds. Composites use the selected
+objective without the old >2-point standalone-relative gate. Their `improvement`
+still means kept win-rate gain over the better standalone on the same covered
+pool, not P&L or baseline-relative gain. See the [4.3.0 behavior changes](../../releases/v4.3.0.md#behavior-changes).
+
+Dataset consumers accept optional `maxMissedSessions` (a nonnegative integer).
+`datasetAge` preserves calendar-day ages and adds missed expected XNYS session ages:
+close columns expect the prior session; open columns expect today's session only
+at/after known-by ET (timestamped open rows without known-by use midnight for this
+daily-session diagnostic, not for as-of matching). Friday close → Monday is 0 missed;
+a missing Friday is 1, and holidays do not count. The calendar covers 2022–2030;
+outside it, age is explicitly `unsupported`. With a limit set, stale/unsupported
+values are refused and count as uncovered in field baselines. Without it, matching
+is unchanged. `datasetWarnings` names stale/unsupported/refused counts;
+`enrich_trades` includes session age in provenance and names `datasetRejections`.
+Session summaries include refused candidates; calendar-day summaries count accepted
+attachments only, returning null ages when none were accepted.
+
 ### Strategy Profile Tools
 
 | Tool                      | Description                                                  |

package/dist/{chunk-YRD62LDE.js → chunk-35RVWZGT.js}

@@ -5181,6 +5181,9 @@ async function getTradesPage(blockId, offset, limit) {
 // ../lib/utils/trade-frequency.ts
 var MS_PER_DAY = 1e3 * 60 * 60 * 24;
 
+// ../lib/calculations/xnys-session-calendar.ts
+var MS_PER_DAY2 = 24 * 60 * 60 * 1e3;
+
 // ../lib/db/index.ts
 var DB_NAME = "TradeBlocksDB";
 var DB_VERSION = 7;
@@ -5671,4 +5674,4 @@ export {
   searchTrades,
   getTradesPage
 };
-//# sourceMappingURL=chunk-YRD62LDE.js.map
+//# sourceMappingURL=chunk-35RVWZGT.js.map