Version 1.17
Prepared September 28, 2026. Market publication is pending submission and review.
This maintenance release retains the Standard / Strong Trend technical trading engine and adds optional data workflows. All external-data features are disabled by default.
Optional FRED Macro Filter
A separate Enable_FRED_Macro switch activates the existing macro profile using the customer’s FRED API key. The four series are real yields (DFII10), broad dollar index (DTWEXBGS), nominal yields (DGS10) and breakeven inflation (T10YIE). The default update interval is four hours. The customer input list exposes only Enable_FRED_Macro. Older presets using only the hidden legacy switch must explicitly enable the new switch.
This feature can affect entry acceptance when enabled. With no key or data not ready, the macro check allows the technical engine to continue. It is bypassed in Strategy Tester and has no new freshness guard; users should monitor data availability.
Optional OpenAI-Compatible Macro Confirmation
The customer Inputs again expose Enable_Macro_AI_Fallback, AI_API_URL, AI_Model and AI_API_Key, with clear labels and defaults off/empty. This is a separate optional confirmation for FRED scores 3/4, not FinBERT news analysis. The prompt includes current and previous FRED observations and the directional score. Only a completed first assistant text choice containing YES passes; errors and invalid replies block configured ambiguous candidates. Scores outside 3/4 use the deterministic macro rules. Standard Tester bypass remains in place.
The response reader now selects direct JSON members and rejects misleading nested/metadata content, duplicate selected fields, malformed output and incomplete generation. A separate no-network MT5 self-test covers request construction, authorization headers, YES/NO and errors, and all 16 FRED direction combinations. A later no-trade live MT5 probe verified FRED retrieval and QwenCloud qwen3-30b-a3b-instruct-2507 connectivity/response reading. Trading benefit has not been verified. See the AI model setup guide.
Trading Economics API Inside the EA
Customers enter their own key in EA Inputs > Your Trading Economics API key and enable Trading Economics news. Add https://api.tradingeconomics.com to the MT5 WebRequest allowlist. With no key, no TE requests are made and the row shows NO KEY; the technical trading engine continues normally.
The EA requests the official News API with Authorization-header authentication. Default polling is every 480 minutes, with a configurable local monthly request limit of 450. Request scheduling and counts persist across restart, sharing quota state by key on the same PC. Failed requests count; this is not the provider’s total usage across other apps or PCs.
It requests up to 30 latest articles and keeps the preceding 24 hours. It does not guarantee complete news coverage between polls. News remains observation only. Requests are bypassed in historical Tester runs. Each request may occupy the EA event loop for up to its five-second timeout. Authenticated live API access has not yet been verified.
The external API collector is no longer needed for customer news fetching. FinBERT still uses a local tokenizer helper, now reading the EA’s te_api_news.tsv file without any API key.
FinBERT ONNX Shadow Analysis
The optional EA reader runs the original ProsusAI/finbert model through MT5 ONNX on English financial text. Tokenization runs outside MT5; the EA records positive, negative and neutral probabilities. Model and token identity, tensor shapes and timestamps are checked.
This is shadow analysis only. It does not block trades, set BUY/SELL direction or change SL/TP. No daily gold-direction score is implemented. The original high-impact news blackout remains separate.
The verified FP32 model is approximately 418 MiB and requires a separate local asset bundle. On the development machine, inference took approximately 0.77-1.20 seconds per article. CPU inference runs in the EA event loop; monitor responsiveness on the target VPS before enabling it.
Dashboard Data Status
The dashboard now shows FRED Macro, the full Trading Economics provider name and FinBERT Shadow. FRED uses actual data readiness; news shows reader status and article count. Hover over each status for details. READY does not imply a trading benefit, complete news coverage or guaranteed freshness.
FRED credentials go in EA Inputs. Trading Economics credentials now go in EA Inputs, like FRED credentials. The panel remains 500 x 585. Trading Economics and FinBERT now share a compact row. Reduced section spacing restores log height from 44 to 134 pixels. It starts expanded and no longer minimizes automatically when chart height is small. Manual minimize/restore still works; a smaller chart can clip the lower controls. The minimized title remains below the top-left price block. Ten no-trade MT5 layout/control checks passed with zero failures.
Maintenance and Validation
- FRED reader handles missing observations and requires two usable values per series.
- Collector credentials are not written into its output or configuration files.
- Mock data is rejected by the production news readers.
- Current external news snapshots are not consumed in historical Strategy Tester runs.
- After independent review fixes: 149 Macro/FRED, 28 TE, 10 dashboard and 22 FinBERT MT5 checks passed; 23 Python collector/tokenizer handoff tests passed. These are module checks, not a full strategy backtest.
- FinBERT MT5/Python ONNX parity was verified on eight synthetic texts within 1e-5 tolerance.
- Version raised from 1.16 to 1.17. Entry, sizing and exit function bodies remain unchanged; initialization and timer add display updates only.
- Customer inputs total 33, including four restored OpenAI-compatible confirmation settings. The legacy FRED switch, research logging, support links and fixed transport paths remain hidden. Research journal and future-bar labels are fixed off.
- The direct API writes Common/Files/AlJamalGold/te_api_news.tsv; the optional tokenizer reads it.
- No trained LightGBM model or new ML trading threshold is deployed.
Read the user manual | AI models and setup | View the audited baseline
