What should Drishti ingest next — and which features actually earn their keep?
June 2026 deep research across four lanes (new data sources, CoinGlass utilization, literature, competitor landscape), a live-probe verdict on switching the analyst model to Claude Fable 5, and a three-tier calibration of the ~100 features the analyst sees every 15 minutes.
Headline verdicts
Funding rate is the only peer-reviewed 24h predictor
Crypto-carry literature (BIS WP 1087) backs funding as a predictive family at drishti's horizon. OI/liquidation cascade lore has no published effect sizes — keep it folklore-graded.
Fable 5: not now
Blocked by the OpenRouter account data policy, ~2.6–4× the cost of Opus 4.8 for this workload (≈$95–125/day vs the $100/day cap), and the current temperature=0 payload is rejected. Full rollout plan in the report if conditions change.
Best new data is free or already paid for
Santiment (30-day lag), CryptoQuant ($109/mo, daily), CryptoPanic ($199/mo) all fail the bar. The shortlist runs on Binance/Deribit free endpoints and the CoinGlass plan drishti already pays for.
A live forward record beats published backtests
Leakage-controlled re-evaluations (FINSABER, Profit Mirage) erase most published LLM-agent wins. Drishti's forward-only track record is the stronger evidence class — the imports worth taking are validation practices, not architectures.
Add-next shortlist (top 6 of 10)
| # | Feature | Cost | Effort |
|---|---|---|---|
| 1 | Spot vs perp CVD divergence | free (Binance WS) | M |
| 2 | IV term structure + IV−RV premium | free (Deribit, wired) | S |
| 3 | Spot–perp basis | free / CoinGlass | S |
| 4 | Global long/short ratio + OI by exchange | CoinGlass (paid already) | S |
| 5 | Coinbase Premium Index | CoinGlass (paid already) | S |
| 6 | Time-of-day priors into the prompt | free (already computed) | S |
Full ranked list with rejection rationale in the report.