Drishti deep research — data sources, Fable, CoinGlass, feature calibration

2026-06-13. Sources: multi-agent web research (23 primary sources, adversarially spot-verified), live API probes with drishti's own keys, the claude-api reference, and a calibration run over 214 live closed signals (snapshot commit c9b7ebc0).

1. Executive summary

• Best new features are free or already paid for. The highest-expected-value additions (spot-vs-perp CVD divergence, basis, IV term structure, global long/short ratio, Coinbase premium) come from Binance/Deribit WebSocket+REST (free) or the CoinGlass plan already being paid for. No new paid subscription clears the value bar — Santiment, CryptoQuant and CryptoPanic all fail on price or staleness (§3).
• Do not switch the live model to Fable today. It is hard-blocked by the OpenRouter account's data policy, costs ~2.6–4× Opus 4.8 for this workload, and the current request payload (temperature=0) is incompatible. Full verdict + a rollout plan if conditions change in §5.
• The literature is on drishti's side methodologically. Published LLM-agent backtests (FinMem/FinAgent/TradingAgents) largely do not survive leakage-controlled re-evaluation; a live forward track record is the stronger evidence class. The actionable imports are validation practices (PSR/MinTRL, overlap-aware CIs), not architectures (§6).
• OI change is the one signal every tier agrees on. Across 214 closed signals, a falling/low 24h OI change predicts wins and a rising/high one predicts losses — it tops the value-bucket spread (76% WR low vs 12% high), tops permutation importance, and is the post-close judge's #1 explanatory feature (oi_change_1h, 11W/1L). Notably this is the family the academic literature calls "folklore-grade" (§6) — drishti's own data disagrees. The analyst already cites OI heavily; this says keep weighting it (§7).
• The statistical layer is underpowered — read it as hypotheses. Only 100 of 214 signals carry a full feature snapshot (the table started mid-stream), leaving 64 won/lost for the regression against 65 features. CV-AUC ≈ 0.80 looks predictive but at p≈n is at real overfitting risk; the citation + bucket tiers (larger N) are the trustworthy ones (§7).

2. Ranked "add next" shortlist

Score = expected orthogonal value × (1/effort) × (1/cost). Effort S < 1 day, M ≈ 1–3 days, L > 3 days, against the existing ~100-feature set.

Explicitly rejected:

3. Paid-API landscape (verified pricing, 2026-06)

4. CoinGlass action list

Current plan pricing (verified at coinglass.com/pricing): Hobbyist $29/mo, 30 req/min, "80+ endpoints" → Startup $79/mo, 80 req/min, "130+" → Standard $299/mo → Professional $699/mo. The docs do not publish per-endpoint tier gating — which is why an empirical probe with the real key is the only reliable map. scripts/probe-coinglass-tier.py (committed, deployed to the VM) tests 19 unused endpoints: aggregated taker buy/sell (CVD), pair taker vol, global long/short ratio, top position ratio, OI exchange-list + aggregated OHLC, basis, liquidation map + heatmap model1, RSI list, Coinbase premium, borrow rate, BTC ETF flows, exchange balance, options max-pain, spot taker vol, whale index, fear&greed, and the account-level endpoint.

Probe results (current Hobbyist key, scripts/probe-coinglass-tier.py, 2026-06-13):

Verdict: stay on Hobbyist ($29/mo list — the paid rate may be a grandfathered $35; the key works either way). Every orthogonal high-value endpoint in the shortlist — basis (#3), global L/S ratio + OI-by-exchange (#4), Coinbase premium (#5), and the CVD inputs (#1) — is already accessible. The $79 Startup tier buys only the liquidation heatmap/map, RSI list, and whale-index family; RSI drishti computes itself, and the degraded top-3-cluster levels via the existing liquidation_heatmap_* features already cover the heatmap need. Don't upgrade unless heatmap-driven SL/TP placement becomes a priority.

Then:

1. Wire {basis, global L/S ratio, top-position ratio, Coinbase premium, OI-by-exchange} into coinglass_advanced.py (same TTL-cache pattern) — shortlist #1/#3/#4/#5, all confirmed free on the current plan.
2. For aggregated CVD, add the required exchange_list param (e.g. Binance,OKX,Bybit) to the two *-taker-buy-sell-volume calls — they 400 only because the param is missing, not because they're gated.
3. Mind the budget: 30 req/min is shared across the 4 existing families; each new endpoint × 8 assets × 15-min cycles needs the same in-memory TTL cache treatment to stay under the cap.

5. Fable (claude-fable-5) verdict: not now

Facts established by probe + the claude-api reference:

Cost projection (anchored on the now-honest telemetry: true burn ≈ $31/day on Opus 4.8): 2× price × ~1.3 tokenizer ≈ $80/day before thinking tokens; always-on thinking on ~770 decide-calls/day plausibly pushes $95–125/day — at or above the $100/day OpenRouter cap shared with drishti-pulse, for an unproven signal-quality gain.

Rollout plan, if revisited:

1. Account: relax OpenRouter privacy settings (user decision — prompt/ feature data would be retained 30 days by the provider).
2. Code: make temperature conditional on model family in llm_client.decide(); add anthropic/claude-fable-5 (request, dotted, and dated-response forms) to COST_PER_M_BY_MODEL in the same commit (the 2026-06-02..12 Opus 4.8 underpricing incident is the cautionary tale); verify forced tool-calling passes a one-shot probe before flipping prod.
3. Switch: time-sliced cohort, not a parallel shadow arm (the histon shadow burned ~$60/day untracked). Flip OPENROUTER_DEFAULT_MODEL, bump prompt version tag (fresh cohort), run ≥2 weeks / ≥40 closed signals, then compare per-version WR/Brier via the existing stats pipeline.
4. Propagation: model change = major change → capabilities entry, changelog, CLAUDE.md, README in the same commit (propagate-change skill).
5. Cheaper experiment first: before paying 3–4× for Fable, A/B the reasoning depth on Opus 4.8 itself (adaptive thinking via OpenRouter reasoning param, if supported through the Bedrock-BYOK route) — same hypothesis ("more deliberation → better calls") at a fraction of the cost.

6. Literature takeaways (full citations in lane synthesis)

On LLM trading agents:

• TradingAgents' Sharpe 5.6–8.2 claims: 3 stocks, 3 months, no costs, no ablations, likely pretraining overlap (2412.20138).
• FINSABER re-evaluation (20 yrs, 100+ symbols, leakage-controlled): LLM agents underperform buy-and-hold by 0.4–0.9 Sharpe with 2–3× drawdowns; FinMem's flagship TSLA result drops to ≈ buy-and-hold (2505.07078).
• "Profit Mirage" (2510.07920): backtested LLM returns evaporate past the model's knowledge cutoff; only causal-reasoning scaffolds generalize. Drishti's forward-only live record is immune to this failure mode — it is the stronger evidence class. Protect it; don't trade it for backtest-validated architecture imports.

On features at the 24h horizon:

• Funding rate is the only family with peer-reviewed predictive evidence at this horizon (crypto carry: BIS WP 1087; funding forecastable: SSRN 5576424). Drishti already weights it heavily — justified.
• Taker/trade-flow imbalance: explains contemporaneous minute-scale moves (Silantyev, Digital Finance 2019) — not established as a 24h predictor. Ingest CVD as regime context, not as a directional oracle.
• OI/liquidation cascade lore: ubiquitous in practitioner blogs, no peer-reviewed effect sizes found — keep graded as folklore in the prompt.
• Options IV/skew: informative for vol, mixed for direction.

Validation imports for a ~200-signal record (ranked):

1. PSR / Deflated Sharpe + Minimum Track Record Length (Bailey & López de Prado) — answers "is 64% WR luck?" while charging for every prompt-version iteration as a trial.
2. Overlap-aware CIs — 24h signals on correlated assets are not independent draws; cluster by time window before any z-test (the tatkaal A/B learned this the hard way).
3. Volatility-scaled barriers — drishti's TP/SL/expire is already triple-barrier; scaling barrier width by realized vol (FreqAI practice) addresses the v23 finding that oversized targets win ~14%.
4. Purged/embargoed walk-forward for any offline prompt-variant evaluation.
5. Hansen SPA when declaring a prompt version the winner.

7. Feature calibration — which features earn their keep

Infrastructure shipped (src/drishti/feature_calibration.py, 12 unit tests):

• Tier 1 — citation win rates: every feature the LLM listed in features_referenced, win rate when cited + Wilson 95% CI, sorted by CI lower bound; judge view (weekly Sonnet attribution) merged via aggregate_attributions.
• Tier 2 — value buckets: tercile win rates per numeric feature from the full signal_features snapshots ("does high funding_z actually lose?").
• Tier 3 — statistical importance: L2 logistic regression won-vs-lost, stratified CV AUC (with an explicit is-this-better-than-coinflip flag) + permutation importance.

Results (214 closed signals, snapshot c9b7ebc0). Pooled baseline win rate 62.6% (97W / 58L / 59 expired). Per prompt-version: v21 50% (64 resolved), v22 100% but only 9W/0L — too tiny to mean anything, v23 68% (81 resolved), v24 just started (1 resolved). The interactive tables are on the research microsite (/explorer).

Tier 1 — win rate when cited (resolved-only, Wilson 95% CI, ≥5 cites):

The post-close judge (feature_attributions, 146 attributed) independently puts oi_change_1h first — 11W / 1L when named as the explanatory feature.

Tier 2 — win rate by feature value (terciles, n≥10): the standout is oi_change_pct — low tercile 76% WR, high tercile 12% (a 64pp spread). fear_greed_value, oi_notional_24h_delta, change_24h_pct and rsi_1h all show the same shape: extreme-high values lose.

Tier 3 — statistical importance (L2 logistic, 64 won/lost × 65 features): CV-AUC 0.80 ± 0.08, and oi_change_pct is the top permutation-importance feature with a negative coefficient (higher OI change → lower win probability), consistent with Tiers 1–2. Caveat: at 65 features on 64 samples (p≈n) this AUC is at severe overfitting risk — treat the ranking as hypothesis-generating, not proof. Only 100 of 214 signals carry a feature snapshot (the signal_features table started mid-stream, ~v23 era), so Tier 2/3 describe a recent cohort, not the whole record.

The one robust, cross-tier finding: OI change is inversely related to win rate — it leads the bucket spread, the permutation importance, and the LLM-judge attribution, all pointing the same way. Given the analyst already cites OI heavily and the data rewards it, this validates the current weighting (and contradicts the academic "OI is folklore" stance from §6 — drishti's own record is the better evidence here).

Honesty rails baked in: expired excluded from WR denominators, n<10 buckets suppressed, Wilson-lower-bound sorting, fixed caveats block (multiple comparisons, non-independence, citation ≠ causation), and an explicit coin-flip flag on the CV-AUC.

8. Competitor capability matrix (lane 4, condensed)

Dashboard ideas worth stealing regardless of features: the funding/liq heatmap panel (#3 in lane 4 — drishti already has all the data), and per-signal "spot-led vs perp-led" tags once shortlist #1 ships.

9. Recommended sequencing

1. Now (free, S-effort): IV term structure + IV−RV (#2), basis (#3), time-of-day prior into prompt (#6).
2. After CoinGlass probe: wire accessible endpoints (#4, #5, part of #1); decide Startup upgrade with data, not vibes.
3. Next build slot (M): spot-vs-perp CVD ingester (#1) — biggest single orthogonal add; Hyperliquid whale bias (#7).
4. Validation pass: PSR/MinTRL on the live record; overlap-aware CIs in the stats pipeline; volatility-scaled TP/SL experiment (directly attacks the 25%-expiry problem).
5. Calibration loop: run feature_calibration per prompt-version cohort; prune dead features from the prompt (cheaper tokens, less noise) and up-grade the folklore-graded ones only if Tier 2/3 evidence supports them.
6. Model: stay on Opus 4.8; revisit Fable when OpenRouter policy/cost conditions change (§5).