What’s new at Stratmill.
New capabilities, useful improvements, and fixes across the research pipeline, trading, the marketplace and the platform, week by week.
26 weeks shown
No updates match your search. Try another term or choose All.
Bring your own exchange, stricter promotion gates, full market-history backfill
This week: live trading with your own exchange accounts, porting strategies to new venues, stricter promotion evidence, full historical data backfills with a coverage ledger, queued Ask-AI answers via email, and a redesigned, localized public site.
Connect your own exchange accounts and trade live on any supported venue
Users can now add their own exchange API keys from the Account page instead of relying on a single shared operator account. Each connected account runs an isolated live engine with its own order ledger, position tracking and reconciliation, and live execution now works across every supported venue rather than one.
Live trading cockpit redesigned around a few key questions
The live trading page replaced walls of engineering text and mostly-empty tiles with a compact status bar, a plain-language readiness summary, a checklist of what is still missing before going live, and risk and reconciliation shown side by side. The Arm control now only activates when the engine is healthy and no halt is in effect.
Port a verified strategy to another exchange
A strategy can now be cloned onto a different venue with its instrument, bar types and leverage limits remapped to that venue's conventions. The clone re-enters verification and must earn its own backtest, optimisation and review results on the new venue's real prices, fees and funding before it can paper or live trade there.
Tighter evidence required before a strategy is promoted
Each idea's out-of-sample holdout window is now locked once and reused wherever it is judged, trial and holdout results now share one consistent profit-and-loss series instead of two different ones, and cross-validation hard-fails a run with missing walk-forward or holdout evidence instead of letting it through. The risk officer now reviews raw numbers independently of the analyst's writeup, promotion floors were regrouped into validity, economics and quality checks, and every paper or live deployment now carries a signed release record naming the approving human and its limits.
Full-history data backfill, Hyperliquid spot venue and coverage tracking
A new backfill program pulls complete price and funding history for every symbol from its listing date across Binance, Bybit and Hyperliquid, repairing interior gaps left by earlier partial runs, while a new coverage ledger tracks expected-versus-actual range per series. Hyperliquid spot markets are now supported as their own venue with correct pricing and fees, and Hyperliquid mark and oracle prices are now recorded historically instead of only being available live. A verified cache keeps downloaded archives until confirmed on disk, and options chains gained daily expiries with faster concurrent discovery.
Ask-AI queues your question when busy and emails you the answer
When every AI seat is busy, a chat question is no longer lost with a generic busy message. It is queued, answered as soon as capacity frees up, and the asker gets an email with the answer and a link back to the conversation, or a notice if it expires unanswered. The operator is alerted if a queued question waits too long. The assistant also holds itself to a stricter rule of evaluating a strategy only on its own venue's data. A rendering fix keeps long chat tables and charts from stretching the conversation past the screen edge on phones.
Public site redesigned, localized, and a business dashboard added
The public marketing site, including the homepage, workflow explainers, FAQ and blog, was redesigned with shared navigation and footer across all pages plus interactive cost-analysis and workflow examples. The full public site and sign-in/sign-up flow is now published in 39 languages with language-aware URLs. An admin business dashboard now shows paginated signup and revenue history alongside recorded AI-usage insights.
Promote-to-paper, smarter document intake, and trading accuracy improvements
Strategy owners can now promote a run straight to paper trading; research intake reads scanned PDFs and embedded charts; paper trading and backtest accuracy, validation rigor and pipeline dispatch reliability all improved.
Promote a strategy straight to paper trading
Strategy owners can now click Promote on a strategy's detail page to move it straight into paper trading, even when the pipeline itself abandoned the run, as long as the owner still judges it worth watching. The system checks the strategy actually has runnable code, a config and a completed backtest before allowing it, giving a clear reason when it can't; promoted strategies carry a visible 'Manually promoted' badge so nobody mistakes it for a pipeline pass.
Research intake reads scanned PDFs and embedded charts
Research intake got smarter about documents. Scanned PDFs with no text layer are now read with OCR instead of being skipped, and charts or diagrams embedded in a PDF, whether as bitmap images or vector-drawn figures, are automatically pulled out and attached as images so the research agents can actually see them, not just the surrounding text. Chat uploads in the tenant workspace get chart descriptions too.
Themed dialogs replace native browser popups on the strategy page
The strategy detail page's 'Build on this', stop and promote confirmations used the browser's native prompt and alert boxes, which look and behave inconsistently across browsers and can't be styled. They now open a themed in-app dialog that matches the rest of the site, supports keyboard use with Escape and Enter, and works with screen readers.
Paper trading no longer books phantom P&L on restart
Fixed two paper-trading accuracy bugs. A routine restart, such as a deploy or process respawn, used to force-close every open position and reopen it fresh, booking phantom profit or loss and paying an extra round trip in fees each time; positions now carry through restarts untouched. Funding payments credited to a paper account previously never reached the account's equity curve; they now do. Live price bars landing a couple of milliseconds short of their close time are also snapped onto the correct boundary.
Backtests and strategy review made more rigorous and realistic
Backtests and strategy review got more rigorous. A bug that let some strategies see a not-yet-published funding rate is fixed. Results are now checked for statistical significance against prior tests on the same instrument, not just the run that produced them, and moving from paper to live trading now requires meeting explicit performance and evidence thresholds. Portfolio managers reviewing a candidate now see its effect on the existing book. Backtests also charge borrow costs on short equity positions, apply per-venue order latency, and consistently see every price bar stamped at its close time.
Fairer, more resilient pipeline dispatch across user workspaces
The pipeline dispatcher was reworked for fairness. Each user's workspace now gets its own eight-slot queue for coding, verification and backtesting instead of sharing one pool with the platform's background research, so a busy account can't delay another user's runs; a user's own runs are also prioritized ahead of platform work. A new watchdog checks every minute for a stuck or queued user run and automatically resumes it. A separate fix stopped concurrent agent calls from occasionally reading each other's answers, which had wrongly abandoned a handful of strategies on a review of the wrong strategy.
More reliable options and funding-rate data collection
Several data-collection failures that could silently corrupt a backtest were fixed. Deribit options-chain requests now retry transient network errors instead of failing outright. Download failures from market-data providers are no longer mistaken for 'no data available' and used to fill backtests with gaps. Hyperliquid funding-rate history now learns each market's inception date with a single request instead of a lengthy re-download that was getting rate-limited and leaving funding data unavailable.
New global markets, cross-market analytics, richer options and news insights
Stocks now span 15 international markets with a new selector, a cross-market Analytics page launched, options and news coverage grew, and mobile/paper-trading UI bugs were fixed.
Stocks page now covers 15 international equity markets
The Stocks page expanded from US-only to 15 markets — US, Hong Kong, China A-shares, Japan, South Korea, Taiwan, India, UK, Germany, France, Canada, Australia, Brazil, Switzerland and Singapore — each with its own trading calendar and session times. A deep-linkable market selector switches the whole page's universe, overview, movers and quotes, and Market Movers can now filter or merge gainers/losers/most-active across all markets, drawing on each market's full research universe instead of a capped sample.
Cross-market Analytics page: correlation, performance and sector tools
A new Analytics page adds cross-market research tools built on world benchmark series, FRED macro and rates data, and company profiles: a correlation heatmap over configurable universes and time windows (with a note when series are excluded for too few data points), a cross-asset performance table, a performance treemap, and sector rotation views. A live World Benchmarks board of index/FX/commodity/rate tiles sits at the top, and correlation universes now include top Binance and Hyperliquid crypto symbols.
Options coverage grows to 66 underlyings with cross-underlying analytics
Options coverage grew from 18 to 66 underlyings, adding country and region ETFs alongside sector, commodity, rates/credit and FX groups. New cross-underlying analytics report ATM implied volatility, term-structure slope, put/call skew, put/call ratios and max pain for each underlying, letting you compare volatility across markets in one view.
News page adds story clustering and expanded international coverage
The News page gained a Story Clusters & Insights section that groups related articles into stories with weighted sentiment, source counts and measured forward price reactions, shown as a scrollable table plus a sentiment-vs-reaction scatter chart in its own full-width row. News source coverage widened with 19 additional international feeds and regional sentiment buckets for China, Japan, Europe, India and emerging markets mapped to the relevant region ETFs.
Mobile table and layout bugs fixed across several pages
Fixed-width tables on the movers, sector and options pages no longer crush their columns to unreadable slivers on phones — they now scroll horizontally with sensible minimum widths instead. Blog cover and hero images stopped stretching out of their intended 3:2 proportions, and card-title subtitles and status labels now align consistently on small screens instead of wrapping awkwardly under other content.
Paper trading list no longer jumps to top; long strategy names wrap
In Paper Trading, the running strategy list previously jumped back to the top every few seconds during its status polling; it now keeps your scroll position. Long strategy names in the detail panel now wrap onto multiple lines instead of overflowing off the edge of the screen on phones.
Deribit historical data gaps now refill instead of stalling
Missing bars in Deribit perpetual futures history previously triggered endless retries every 20 seconds with no progress. Gap-filling now re-fetches the missing series directly from Deribit's full-history endpoint, closing gaps that used to stay open indefinitely and improving the completeness of data available for backtesting.
Options data history, a new blog, and broader translations
This week: forward-looking US options market data collection began, a daily multilingual blog and a new Data page and Contact page launched, site translations expanded, landing visuals were refreshed, and a paper trading chart bug was fixed.
Historical US options market data collection
Stratmill now continuously captures snapshots of US options chains, including bid/ask, implied volatility, open interest, volume and full greeks, for a defined set of symbols, polling roughly every 30 minutes during market hours. Because usable licensed historical options data isn't otherwise available, this builds an internal point-in-time history from this point forward for future strategy research.
Daily blog with multi-language support
A new blog was added to the public site, publishing an article on quant trading and backtesting topics roughly once per day. Posts and the blog index are translated into over thirty languages, use the same navigation and footer as the rest of the site, and are linked from the site navigation and sitemap.
New public Data page and Contact page
Visitors can now browse a dedicated Data page describing the market data Stratmill collects, and reach the team through a new Contact page with a support channel. Both pages share the same navigation, footer and language switcher as the rest of the public site.
Expanded translations across the site
Legal pages, the data page, blog, and in-app screens for news, options, US stocks and the trading dashboard now have more complete translations across the platform's supported languages, closing gaps where labels or notes previously showed only in English or, in a few places, went blank due to a script conflict that has now been fixed.
Consistent navigation and sign-in state everywhere
The signed-in account menu and Dashboard link now correctly appear on every public page that shares the landing navigation, including all blog pages, the data page and legal pages, instead of only on the homepage. Cached dictionary files were also refreshed so language updates reach visitors reliably instead of showing stale text.
Landing page visual refresh and dashboard usability
The homepage's hero terminal, feature and pricing cards moved to a flatter, borderless look, and the final call-to-action panel now uses a solid accent color for emphasis. In the dashboard, related strategy pages are now grouped under a single collapsible menu section, and the strategies table keeps its filters and header visible while scrolling. Several overview cards that had invisible backgrounds or overflowing text in some languages were also fixed.
Paper trading drawdown chart missing replay line
When switching between strategies on the paper trading page, the dashed backtest-replay overlay could disappear from the drawdown chart while still showing correctly on the equity chart. The chart now clears its reference to the old series when strategies are switched, so the replay line consistently appears on both charts.
Prediction markets, US equities, and broader market intelligence
Polymarket prediction markets gained backtestable pricing with real spreads, US equities became a fully supported market end-to-end, news and market-overview coverage expanded widely, and strategy charts plus paper-trading restarts got more accurate.
Prediction markets now backtestable with real spreads and settlement
Polymarket-based prediction markets can now be backtested as binary options that settle at the resolved outcome price, using real bid/ask spreads replayed from collected order-book quotes instead of an assumed constant. The Prediction Markets page also now shows real market titles instead of raw slugs, plus sortable columns and type/search filters.
Dedicated Market News page with broader sourcing and sentiment
A new Market News page collects headlines from dozens of sources, including Reuters, Bloomberg, WSJ and full Google News section coverage, with real publisher logos, search, and topic/asset filters. Sentiment charts now compute a series for every tagged asset in the news window instead of a fixed set of buckets, and PREDICTION/OPTIONS categories were added to the scoring vocabulary.
US equities: session-aware backtesting and a full live market page
US stocks can now be researched and backtested against a real NYSE session calendar with historical data sourced from Yahoo. The US Stocks live page gained the same depth as the crypto Live Market page: streaming order book and trade tape with whale marks, all standard timeframes, cumulative volume delta, and whole-market top movers.
Market overview cards for crypto, US stocks and options
New overview cards summarize each market at a glance: total crypto market cap, dominance, Fear & Greed and trending coins on the crypto page; indices, VIX, yields, gold, oil and sector performance on the US Stocks page; and aggregate open interest, put/call ratio and 24h volume on the Options Chain page.
OKX and Bybit added as options chain venues
The Options Chain page now pulls live chains from OKX and Bybit alongside the existing venue, adding coverage for BTC, ETH and SOL options. The chain, expiry-overview and most-active tables also switched to fixed-header, independently scrollable panels so long tables no longer distort the page layout.
Strategy chart now always shows the strategy's own bar timeframe
The Price & Trades chart on strategy detail pages could previously substitute a coarser timeframe, such as daily candles for a 2-hour strategy, to keep large payloads small, which risked misrepresenting intraday entries. It now always renders native bars, with older history lazy-loaded as you scroll left, removing any risk of a misleading chart.
Paper trading positions now carry through restarts without resetting cost basis
Restarting a paper trading node used to book a synthetic close-then-reopen pair for every open position, resetting its cost basis and fragmenting the trade history. Positions now carry through restarts at their original entry price, so unrealized P&L and the eventual close both read as one continuous trade across any number of restarts.
Paper trading analytics overhaul and pipeline reliability guardrails
Paper trading gains a full analytics dashboard with a backtest-replay overlay, plus fixes to crashes and PnL math. Pipeline guardrails prevent false failures and false kills. Data coverage reporting and COIN-M backtest accounting are corrected.
Paper trading gets a full analytics dashboard with backtest-replay overlay
The paper trading page now shows side-aware stats (long/short counts, win rates, average win/loss, streaks), long-vs-short and fee-drag charts, and a per-trade return chart, grouped into labelled sections. Restart bookkeeping rows are separated from real fills so trade counts reconcile. A dashed line now overlays the daily backtest replay's equity and drawdown curves on the live paper curves, showing whether paper trading is tracking what the backtest predicted.
Paper trading page crashes and PnL percentage errors corrected
Opening the strategy-detail panel on the paper page could error out due to an immutable-row bug when pairing entries with exits; that is fixed. Per-trade PnL percent was computed with the wrong long/short formula, misreporting short-trade returns; it now uses the correct side-aware entry notional. A bug that could blank out later charts when an earlier one had no data was also fixed.
Data coverage reporting now reflects what is actually obtainable upstream
The data coverage view adds an Obtainable column next to Avg Coverage, counting gaps that have been verified absent from upstream sources as complete rather than as missing. A related fix corrected supplementary kline gap refills, which had been requesting the wrong candle interval and wrongly marking over a thousand rows as unfillable.
Guardrails stop the research pipeline from recording false failures
Several safeguards now catch runs that get stuck or misjudged: a run stalled at the same stage while still billing tokens now pauses instead of burning a full budget; crashing after a paid agent call no longer wipes out progress or starves other runs in the same stage; malformed agent output no longer crashes handlers; NaN results now fail checks instead of passing them silently; and a strategy that produces zero trades on data too thin to judge is now parked and re-checked rather than recorded as a failed idea.
Watchdog and timeout tuning stops healthy runs from being killed early
A liveness watchdog was firing on healthy pipeline passes that were simply processing several runs in sequence, restarting the whole factory unnecessarily; it now tracks progress per run instead. Backtest heartbeat tolerance was raised and made configurable so shared-host load spikes stop killing live backtests, and the developer agent now gets a longer timeout to match its heavier workload, cutting spurious verification failures.
COIN-M inverse-contract accounting corrected in backtests
Market impact costs and funding accrual for inverse COIN-M contracts were computed in the wrong currency, multiplying costs by the coin price and making every COIN-M trade look like a loser. Both are now charged in the correct settlement currency, so COIN-M strategy backtests report accurate win rates and PnL.
Optimizer respects declared parameter bounds; parity scoring uses a rolling window
Strategies can now declare hard parameter bounds so the optimizer only searches values the generated code will actually run, instead of testing values that get silently clamped and reported as if they ran unmodified. Separately, the daily paper-vs-backtest parity check now scores only a recent rolling window, so a single bad week or a restart no longer keeps a strategy permanently flagged for review.
Live Trading Execution Engine, Paper Parity Tracking, and Data Fixes
Live trading gained a full execution engine with real drawdown limits and mirror-integrity checks, paper trading now shows daily parity vs backtests, inverse-contract cost bugs were fixed, and the Ask-AI chat dock got layout fixes.
Live trading execution engine with drawdown protection and mirror-integrity checks
A live execution engine has been built that mirrors each paper strategy's fills onto the owner's exchange account at equity-scaled size. Every order passes dedup, staleness and risk checks before landing in a ledger. Real drawdown limits are enforced from a durable equity high-water mark, and a new integrity check halts trading if a live position diverges from what paper already exited. The cockpit now shows account equity, drawdown vs high-water mark, and this mirror-truth status.
Paper trading page: daily parity tracking and faster browsing
A new daily table compares each graduated strategy's paper trading decisions against a replay of its backtest, showing matched vs. mismatched trades and entry/exit timing drift with an OK or REVIEW status per strategy. The strategy list also now opens on the first strategy automatically instead of an empty pane, and stays visible while scrolling a long list.
New backtest data: per-leg open interest and order-book imbalance
Cross-sectional and basket strategies can now use open interest tracked separately for each leg of a trade, the same way funding rates already are. A new order-book-imbalance signal, derived from resting bid/ask depth data, is also available as backtest input.
Inverse-contract backtests no longer overstate trading costs
Backtests on inverse contracts such as ETHUSD_PERP were computing trade notional incorrectly, inflating simulated market-impact costs by roughly 200x on some pairs and driving those backtests into forced liquidation. Notional is now calculated from the correct fixed contract size, restoring realistic cost estimates.
Hyperliquid funding-rate collection fixed
The Hyperliquid funding-rate collector was calling the exchange once per symbol and using a symbol-matching rule that never matched Hyperliquid's naming convention, so it silently collected no funding data while hitting rate limits. It now fetches the whole market in a single call and matches coins correctly, including special-cased symbols.
Ask-AI chat dock: layout and reliability fixes
Wide tables and charts could overflow the small assistant dock and spill past its edges; both now scroll within their own frame, with charts kept at full readable size instead of shrinking. A stray arrow artifact in the thinking spinner was removed, and responses still streaming when a user switches pages or sections no longer disappear — they resume from where they left off.
Faster symbol search on the Data page
Typing in the Data page's symbol filter used to scan over a million database rows per keystroke, often taking over three seconds before any options appeared. Symbol matching now runs against a background-refreshed in-memory list, cutting typical response time from seconds to a few milliseconds.
Live Market Expands to Four Exchanges, New Options Chain, and Trading Safety Groundwork
Live Market now covers four exchanges, a new Options Chain page shipped, public pages load faster, a paper-trading scheduling bug and a backtest chart bug were fixed, and early live-trading safety groundwork was laid (no order execution yet).
Live Market page now covers Binance COIN-M, Hyperliquid, Bybit and Deribit
Live Market expanded from Binance USD-M futures only to also include Binance COIN-M perpetuals, Hyperliquid, Bybit (linear, inverse and spot), and Deribit perpetuals. The single flat market picker was replaced with a two-level Exchange then Market selector, so it is clear which exchange a symbol's live price and order-book data are coming from before it is used.
New Options Chain page for Deribit and Binance European options
Added an Options Chain page covering Deribit (BTC, ETH, SOL, XRP and other USDC-settled underlyings) and Binance European options: a calls/puts table with bid, ask, open interest and implied volatility, an IV smile chart, an ATM IV term structure chart, open interest by strike, a greeks toggle, and expiry/most-active summary tables. Early bugs were fixed where put bid/ask columns were swapped, altcoin chains loaded empty, and stat tiles overlapped their neighbors.
Public pages load faster and no longer show the dashboard shell to visitors
Anonymous visitors opening Live Market or Options Chain from the public landing page now see a content-only view instead of the full signed-in dashboard shell with app navigation links that all bounced to login. Separately, pages that previously loaded fonts and JS libraries from third-party CDNs now self-host them, so first paint no longer stalls or blanks out when those CDNs are slow or unreachable.
Multi-leg paper strategies no longer run on a permanently offset schedule
Multi-leg paper trading strategies were fed warmup bars one leg at a time (all of one asset's history, then the next), so a strategy reading cross-leg state during warmup saw other legs empty and only began rebalancing after its first live bar, offsetting its schedule from the matching backtest by days. Warmup bars are now interleaved chronologically across legs to match backtest order.
Backtest charts no longer show phantom trades after a liquidation
A backtest that got liquidated stopped updating its equity curve at the bust point but kept generating trade markers through the rest of the data, so the price chart could show years of trades on an account that was already wiped out; trade lists and chart markers now stop at the liquidation point. Marketplace subscription copies, which stay parked and never move stages, were mislabeled 'Backtest' and now read 'Marketplace Copy'; the Publish to Marketplace button also got clearer styling.
Agents operations page and quota failover reliability fixes
The Agents operations page read pause and error state from the wrong key and always reported every agent as active with zero errors, even during multi-hour quota pauses; it now reflects real pause and failure state. Separately, a bug in how quota-capped accounts were cooled down caused agents to keep retrying one capped account instead of switching to healthy ones, contributing to an extended agent outage; cooling now works correctly.
Early groundwork laid for live trading: safety gate, risk gate, and cockpit page
Began building the infrastructure needed before any account can trade live: a fail-safe trading gate that defaults to disarmed and treats a missing or unreadable safety state as halted, a pre-trade risk gate, and position reconciliation, plus a live-trading cockpit page and API contract for the owner to review. This step intentionally does not yet let the system place or cancel any live order.
Weekly Update: Funding-Accurate Paper Trading, SaaS Launch, and Pipeline Quality Gates
This week: paper trading now matches backtest funding and leverage, the public SaaS app and a second payment provider launched, assistant chat streaming and attachments improved, and the research pipeline enforces quality gates with fewer stalls.
Paper trading now mirrors backtest funding and leverage exactly
Before, funding payments and leverage weren't reflected in the paper trading sandbox account, so a strategy's paper equity curve missed funding income until a position closed, and leveraged strategies silently ran at 1x margin. Now perp funding accrues directly into the sandbox account each funding window, using live rates with a fallback source when the database lags, and paper venues are margined at the same clamped leverage the backtest used. A daily job also compares each strategy's paper trades against a fresh backtest replay to catch drift, and restart-related trades are now timestamped and labeled accurately.
Fixed paper trading chart corruption and color glitches
Duplicate-timestamp equity snapshots from restart reconciliation could break the charts' ascending-time assumption, producing a full-width horizontal streak across the equity and drawdown views; charts could also render blank when data loaded late. The same corruption affected the kline and signal charts. All four now normalize and deduplicate data before rendering, rebuild correctly after page recovery, and use the platform's standard blue color instead of a legacy purple.
Assistant chat gets faster streaming, working attachments, and safer output
The in-app assistant chat now starts streaming replies immediately instead of buffering the first response, auto-generates a short conversation title after your first message, and lets you stop a reply mid-stream. File attachments are now actually read and used in answers instead of being ignored. Chat credentials rotate across a shared pool so busy periods fail over instead of stalling, and internal operational notices, such as rate-limit messages, are filtered so they never appear as garbled text in a chat bubble.
Stratmill launches as an installable, multi-provider SaaS app
The full public-facing web app, including landing, login, account, marketplace and pricing pages with multi-language support, plus its backend for accounts, billing and credits, is now live and version-controlled. The site can be installed as an app on phones and desktops with an offline fallback. Payments now support a second checkout provider alongside the existing one, and billing edge cases including double-charging on plan upgrades and a marketplace purchase race are fixed.
Fixed landing page layout breaking on small phone screens
The pricing comparison table, fees table, and feature-rigor grid on the landing page overflowed or clipped content on narrow phone screens around 390px wide, hiding columns instead of letting users scroll to see them. All three now fit and scroll correctly on mobile devices.
Strategy research pipeline now enforces quality gates and learns from failures
Quality checks such as overfitting flags, out-of-sample performance and statistical significance were previously advisory only, so a strategy that failed them could still be promoted; hard and soft quality floors are now enforced before promotion. Failed and abandoned strategy attempts are now recorded into a searchable history so research avoids repeating near-duplicate failed ideas. A capacity bug that let waiting strategies block all new research for weeks is fixed, along with several correctness fixes to the underlying decision and task-tracking logic.
Fixed gaps in historical market data parsing
Two parsing bugs in the historical data importer were dropping information: fractional order-book depth levels near the price were being rounded away, losing about one in twelve levels per snapshot, and the earliest archived data files from 2022-2023 were being skipped entirely due to a date-format mismatch. Both are now parsed correctly, improving the completeness of historical data used for research and backtesting.
COIN-M inverse futures support and paper-trading fixes
This week added support for trading Binance COIN-M inverse perpetual futures in paper trading and backtesting with correct accounting, inverse-aware position sizing, and a Hyperliquid paper-trading fill-timing fix.
Paper trading now supports Binance COIN-M (inverse) perpetual futures
Binance's COIN-M inverse perpetuals previously could not be traded in paper mode because the internal symbol naming didn't match what Binance's market data feeds expect, causing the connection to fail to start or receive no price updates. The symbol handling has been corrected so COIN-M instruments load properly and receive live quote and bar data alongside existing USD-M and spot markets.
Corrected notional, fee and PnL accounting for COIN-M contracts
COIN-M inverse contracts were being priced as if they were standard linear contracts, producing wildly inflated notional values and fees. Paper trading and backtesting now treat these instruments as inverse contracts with the correct per-contract USD value, so position sizing, fees, and PnL for COIN-M strategies are accurate and consistent between backtests and paper trading.
Strategy position sizing now accounts for inverse-contract instruments
Strategies trading inverse contracts (like COIN-M) previously used a sizing formula intended for linear instruments, which could mis-size positions by large factors or let orders below the intended minimum size through. Strategies now compute minimum order size and position size correctly for inverse instruments, sizing in whole contracts instead of misapplying linear coin-unit math.
Fixed early time-stop exits on Hyperliquid paper trades
On Hyperliquid, paper-traded positions could record their open time slightly later than the bar they were actually entered on, causing time-based stop-loss exits to trigger one bar later than they would in a backtest of the same strategy. Fill timestamps are now aligned to the exact bar the entry decision was made on, so time-stop behavior matches backtest results.
TradingView charts, paper trading upgrades, and execution fixes
Strategy detail pages gained interactive TradingView-style charts, paper trading got performance/funding/multi-leg tracking plus Hyperliquid support, order-fill parity and reliability fixes landed, and data-coverage/page-caching were cleaned up.
Strategy detail pages get interactive price and performance charts
Backtest results now render with TradingView-style interactive charts: equity curve, drawdown, cumulative return, rolling Sharpe, and options greeks/IV, plus a price chart overlaying trade entry and exit markers on candles. Multi-instrument strategies get a leg selector, options strategies show trade markers on the underlying price, and multi-timeframe single-instrument strategies are now supported instead of the chart being hidden.
Paper trading dashboard adds performance metrics, funding, and multi-leg positions
The paper trading detail view now shows performance cards (return, win rate, profit factor, average trade, expectancy, max drawdown, Sharpe/Sortino), reconstructs funding cost per closed trade the same way backtests do, and displays options trades with strike/expiry columns and USD-valued premiums. Basket and pair strategies now track every open leg instead of only one, and Hyperliquid perpetuals are now supported for paper trading.
Improved paper trading and portfolio tables
Trades tables on paper trading and strategy detail pages now support filtering by instrument (labelled with market type, e.g. Perpetual vs Spot) and pagination for large trade histories. The hypothesis tag list collapses to one line with a '+N more' expander. The portfolio page now shows a compounded equity-curve chart, a stage column for each strategy, and a data freshness timestamp.
Fixed paper trading order fills and restart reliability
Paper trading market orders previously filled only at the size of the coincident live trade tick, leaving positions far below the intended size; orders now fill fully like in backtests using a real order-book feed, and fill timestamps are snapped to the same bar grid backtests use so time-based exits trigger on the correct bar. A strategy promoted while the trading node was running no longer crashes the node if its instrument wasn't already loaded. Spot strategies can no longer open a short position when a sell order exceeds the held balance.
Fixed funding accrual and options strategy verification
Backtests now correctly accrue funding on perpetual positions that are still open when the test period ends, instead of recording zero funding despite being held through settlements. Automated verification also no longer rejects options strategies whose option legs are only generated at backtest time; those legs are now expanded before the verification run, so options strategies can pass verification.
Cleaned up the data coverage table
Hyperliquid spot symbols in the data coverage view no longer show the internal spot_ prefix (e.g. PURRUSD instead of spot_PURRUSD), and the unused Refill column was removed to simplify the table.
Fixed stale page caching and a scrollbar rendering gap
Pages were sometimes served without cache-control headers, so UI updates could fail to appear without a hard refresh; HTML pages now send no-cache the same way JS/CSS files already did. Also fixed a visible gap next to the scrollbar in trades table headers, most noticeable in light theme.
Weekly Update: Self-Heal Loop, Wider Cost Tracking, and a Full Data Catalog Browser
This week added an autonomous self-heal code-fix loop with a review dashboard, expanded AI cost tracking with new usage charts, overhauled the Data page into a full dataset browser, and shipped several UI consistency and reliability fixes.
Autonomous self-heal loop investigates and drafts fixes for backtest bugs
When a backtest engine inconsistency persists after re-running with fresh data, the platform can now autonomously investigate the root cause, draft a code fix, and verify it against the test suite plus a sample of unaffected strategies before pushing the fix as a branch for review. A new Self-Heal page shows every fix attempt's findings, root-cause analysis, diff, and verification verdict, keeping merges human-gated as the review surface.
AI cost tracking now covers every agent, with new usage charts
Token-cost tracking now covers all AI agents on the platform, including the new self-heal loop, the semantic auditor, the chat assistant, and the Telegram assistant, not just the core research-pipeline agents. The dashboard also gained a usage-over-time chart and a cost-by-agent chart, both with an agent selector and a toggle to switch between cost, call count, and token count.
Data page now browses the full dataset catalog, not just the worst-coverage summary
The Data page previously showed only a worst-coverage summary; it now lists every dataset the platform has collected, with server-side pagination, Exchange and Venue columns with filters, a symbol search box, and readable source names. Every row can be opened to preview its underlying data, and the coverage table now also shows duplicate counts and refill status.
Strategies table replaces four verification columns with one compact Verify column
The Strategies table previously showed four separate verification-layer columns (static, synthetic, sandbox, QA). These are now combined into a single Verify column with small colored pass/fail/skip indicators and tooltips for each layer, freeing up table space for metrics like return and win rate.
Remaining theme mismatches fixed; more dropdowns use the searchable themed menu
A few pages that had fallen back to a mismatched purple accent color and background now match the rest of the site's theme. The last plain dropdown menus on the site — Live Market's symbol, interval, and market pickers, the audit run selector, the portfolio scheme selector, and the dashboard cost-window selector — now use the same searchable, themed menu as the other filters.
Clearer factory health readout and new pipeline-stage timing chart
The factory memory usage bar on the dashboard was invisible on the light theme; it now shows clearly in both themes with a tooltip explaining the reading. A new Time-per-Pipeline-Stage chart shows how long strategies spend in each stage of the research pipeline, such as coding, verifying, backtesting, and optimizing.
Fixed a backend startup crash and sped up historical data re-collection
Fixed a missing import that could crash the web backend on startup under certain conditions. Also sped up historical data collection so re-syncing a symbol now only downloads and catalogs genuinely new dates instead of reprocessing its entire history on every run.
Options analytics, portfolio insights, and automated pipeline auditing
Added options analytics and a portfolio correlation view, introduced hourly pipeline auditing with per-agent cost tracking, and improved backtest accuracy with mark-to-market equity curves, robustness metrics, and funding fixes.
Options-specific analytics on the strategy detail page
Strategies built on options now show a dedicated analytics view instead of the generic layout: net greeks (delta, theta, vega) at entry, roll win rate, percent of contracts expired worthless, implied versus realized volatility, the leg structure, and a per-roll schedule table. Spot and futures strategy pages are unchanged.
New Portfolio view with correlation and risk-contribution breakdown
A new Portfolio page shows the graduated strategy book's overall Sharpe versus a weighted average, a diversification ratio, effective number of independent bets, and a correlation heatmap between live strategies. Candidate strategies that are highly correlated with the existing book are now factored into promotion review so the book doesn't concentrate risk in similar strategies.
Hourly pipeline auditor and per-agent cost tracking
A new Audit page reports hourly health checks across the pipeline, covering data-integrity issues on individual strategies, stuck or failing agents, and infrastructure problems, with a run-history selector to browse past audit runs. The dashboard also gained a Token Cost card breaking down each agent's calls, tokens, and spend over selectable time windows, plus pagination for the Recent Activity feed.
More rigorous risk and robustness metrics on strategy results
Backtest equity curves are now marked to market, catching intra-trade liquidations when leveraged positions hit realistic maintenance-margin levels rather than only reflecting realized PnL at exit. Strategy detail pages now also surface tail-risk stats (VaR, CVaR, Omega, tail ratio), benchmark-relative alpha, beta and information ratio, deflated/probabilistic Sharpe ratios that correct for testing many parameter combinations, and a leverage badge showing effective exposure per trade.
Corrected options equity curves and funding calculations
Fixed native options backtests showing wildly inflated equity curves caused by settlement collateral being valued at spot price, and fixed exercised or expired option trades showing zero PnL despite a nonzero return. Funding payments are now valued against the real mark price instead of the bar close, and COIN-M perpetual funding collection, which previously failed outright, now works correctly.
Broader Deribit options data, rolling-options strategies, and accurate funding costs
Expanded Bybit and Deribit market data, added rolling options strategies with trade-level greeks, made backtests account for real funding costs, and improved Data Sources UI plus several reliability fixes.
Bybit on-demand collection and full Deribit data coverage
On-demand data collection now covers Bybit (previously failing with 'no collector available' errors) and the full range of Deribit instruments: spot, perpetuals, dated futures and options. Backtests can also pull Deribit's forward-looking DVOL implied-volatility index and perpetual funding-rate history as signals.
Rolling options strategies with per-trade greeks
Options strategies can now be defined as recurring structures, such as weekly short strangles or covered calls, that automatically roll into new contracts as they expire; the full expanded contract chain is shown in the strategy's Data Sources. Each options trade now records its strike and type, plus entry greeks when available, rolled up into a strategy-level options summary.
Perpetual funding now credited/debited in backtests
Backtests now reconstruct the funding payments a perpetual futures position actually pays or receives at each funding interval and fold them into the account balance, rather than ignoring this cost/income. This now applies to every linear-perpetual position, not just strategies that explicitly read funding data, and multi-leg baskets accrue each leg's own instrument funding rate instead of reusing one leg's rate for all.
Data Sources and trades table improvements
The trades table now shows a separate Price PnL column alongside funding and net PnL. The Data Sources panel surfaces all supplementary and subscription data a strategy uses (including Deribit greeks/IV and funding feeds), correctly labels the market for Deribit, COIN-M and options instruments instead of 'Unknown', shows accrued funding for perpetual strategies, and displays long tables in a scrollable, bordered box matching the trades view.
Reliability fixes across agents and dashboard
Fixed a bug where large trade histories made the summary agent's prompt exceed length limits, causing repeated failures to generate strategy summaries. Added a retry to the database health check that was raising false 'unreachable' alerts under load, sped up data-collection instrument lookups that were freezing on large catalogs, kept multi-word status badges on one line, and completed the strategy stage filter with all pipeline stages.
Paper trading now waits for all legs before signaling
Multi-instrument paper-trading strategies could fire an entry or exit signal before a slower-arriving secondary leg's bar had been received, trading on a stale price for that leg. Paper trading now waits, within a bounded window, for every same-timeframe leg to deliver its bar before evaluating the signal, matching the equivalent safeguard already used in backtests.
Paper Trading Hyperliquid Support, Strategy Outcome Summaries, and Data Fixes
This week added Hyperliquid paper-trading support, a new strategy outcome-summary feature, a hypothesis-revision salvage path, and a cluster of fixes to strategy-page displays and market-data collection reliability.
Paper trading now supports Hyperliquid legs, in-signal order strategies, and a position risk cap
Paper-trading workers previously only worked with Binance market data and a single order-submission style, so Hyperliquid legs and strategies that submit orders directly inside their signal logic silently failed to trade. Paper trading now streams Hyperliquid market data for extra legs, executes both order-submission styles, and caps any single position at a configurable share of buying power to guard against oversized fills.
New Outcome Summary card explains how each strategy's journey ended
A new summary agent generates a short, factual account of what a strategy was, how it performed, and why it reached its final pipeline stage. The summary is produced at key milestones (abandon, promotion, paper, live, retirement) and is now shown on the strategy detail page as an Outcome Summary card.
Flawed hypotheses can now be reframed instead of only iterated or abandoned
Previously a hypothesis with no real edge could only be re-coded on the same dead premise or abandoned outright. A new revise-hypothesis path sends salvageable-but-flawed ideas back to research for reframing (different instrument, timeframe, threshold or mechanism) before giving up, and the strategy page now shows this outcome plus how many times an idea has been revised.
Strategy detail page shows more accurate and readable pipeline history
The strategy detail page now surfaces the risk officer's review, shows each iteration's own backtest result and true failure reason instead of a stale run-level one, presents a calmer collapsible abandon-reason banner, and formats hypothesis and analysis text into labeled, readable sections.
Fixed several causes of strategies being wrongly abandoned for missing data
Corrected routing so COIN-M futures data is fetched from its correct source, fixed collection of extra-instrument legs in pairs and cross-venue strategies, wired liquidation events and Hyperliquid predicted-funding data into automatic collection, capped an oversized tick-data collection window that could crash the pipeline, and stopped endlessly retrying data windows that genuinely don't exist upstream.
Developer agent gets clearer failure diagnostics and its own prior code
Verification feedback now pinpoints exactly why a strategy produced no trades (entry never triggered, sizing rejected, or below minimum notional), and the developer agent is now given its previous code version on each iteration instead of re-deriving from scratch, reducing repeated regressions and back-and-forth churn. Auditor memory usage was also flattened to avoid mid-run restarts.
Paper-trading signal history and chat session management added
Added a signal-history view for paper-trading strategies and a chat session sidebar with the backend endpoints to list and delete sessions, giving a clearer record of paper-trading activity and prior conversations.
Faster pipeline dispatch, clearer trade venue columns, and data-loading fixes
This week's changes focus on internal platform reliability: faster and fairer strategy processing, plus fixes to the trades table display and a data-preview loading bug.
Coding stage now dispatches strategies in parallel and processes backlogs fairly
The strategy coding stage previously processed strategies one at a time and could repeatedly re-pick the same batch, leaving a large backlog of waiting strategies stuck until they timed out untouched. Coding runs are now dispatched concurrently (up to 6 at a time, up from 3), the queue processes oldest-waiting strategies first, and a strategy that has never actually been picked up by the agent is no longer timed out and abandoned. Together these changes mean large backlogs of strategies move through coding faster and more predictably instead of stalling.
Temporary quota pauses no longer repeatedly overshoot into multi-hour outages
When the coding agent's usage quota was temporarily exhausted, the system treated it like a run of ordinary failures and escalated to increasingly long cooldowns, so processing could stay paused well after the quota had actually reset. Pauses now end at the actual subscription reset time instead of following the generic cooldown schedule, avoiding needless extended downtime.
Audit cycle no longer stalls indefinitely on a single hung network task
The background auditor process could appear alive but stop making progress entirely because one hung network or disk task would block the whole cycle from completing. Per-task and overall time limits now force slow or stuck tasks to fail and release their slot so the audit cycle keeps moving instead of freezing.
Large trade-data downloads no longer risk crashing the collection process
Requesting trade-level data for a strategy could trigger a download spanning years of history loaded entirely into memory at once, risking a crash on high-volume symbols. Downloads are now streamed and written incrementally, and the requested window is capped to the last 30 days, keeping memory use bounded regardless of how much history was requested.
Gap-filling data requests now try the direct connection before proxies
Refilling missing market data previously always attempted proxy connections first before trying a direct connection, wasting time on hosts that could reach the data source directly. The direct attempt is now tried first, falling back to proxies only when it fails, speeding up gap-filling on hosts with direct access.
Data preview page no longer hangs on a spinner after a failed first load
The data preview page's chart data was being decoded with an incorrect binary format, causing every request to fail; on a first-time page load, that error had nowhere to display so the loading spinner stayed on screen indefinitely. The backend now decodes bar data using the correct format, and the frontend properly displays load errors so the spinner clears even on a first-load failure.
Trades table now shows exchange and instrument type as separate columns
The strategy trades table previously showed only a symbol, making it hard to tell which exchange and instrument type (spot, perpetual, futures, or option) executed a given trade, especially for multi-venue strategies. The table now splits this into distinct Symbol, Exchange, and Type columns for clearer per-trade detail.
Weekly Update: Data Pipeline Reliability Fixes and UI Polish
Data pipeline reliability fixes (auditor stalls, Bybit rate limits, missing weekly/monthly and COIN-M history) plus two frontend fixes: strategies list pagination and full-precision trade prices.
Restored stalled market-data auditing and backfill
Several regressions had crept into the automated data auditor: a query-ordering bug meant duplicate-data cleanup never ran while gaps kept growing, a catalog-write mismatch silently killed every refill, and a missing liveness heartbeat caused backtests to be rejected as running on stale data. A blocking database write on the auditor's main loop was also causing spurious connection timeouts during refills. All are fixed, restoring gap-filling, duplicate cleanup, and reliable backtest data-freshness checks.
Bybit data refills throttled to avoid rate-limit errors
Bulk historical data refills against Bybit were bursting past its API rate limit, producing large batches of failed refill attempts with unhelpful blank error messages. Refills to Bybit now run through a dedicated, lower concurrency limit while other exchange data sources keep running at full speed, and failure messages now include the underlying error type to make problems easier to diagnose.
Auditor memory usage now tunable per deployment
The background data-auditor process used fixed concurrency settings that could exhaust memory and interrupt data-quality checks on smaller servers. Parallelism and batch sizes are now configurable per deployment, letting the process run safely within a given server's memory budget without code changes.
Weekly and monthly price history now stays up to date
3-day, weekly and monthly candle history stopped updating after the upstream archive that fed it was discontinued last year, so lower-frequency symbols were missing recent data. These intervals are now fetched directly from the exchange API, so historical price data for strategies using longer timeframes stays current.
Fixed missing COIN-M futures price downloads
Historical candle downloads for COIN-margined futures contracts were silently returning no data because the download routine used the wrong data path. Plain candle downloads for these contracts now resolve correctly, matching the related price feeds that already worked.
Strategies list keeps your page after a refresh
Reloading the strategies page while browsing past the first page reset the view back to page one, losing your place in a long strategy list. The page position is now remembered and restored on reload.
Trade prices now display at full precision
The trade history table always rounded entry and exit prices to two decimal places, which hid meaningful digits for lower-priced assets. Prices now display with their full native precision, while the underlying backtest and optimizer calculations are unaffected since they already used full precision internally.
Leverage support, expanded venue routing, and data-pipeline reliability fixes
This week added leverage and seven-venue routing to strategy backtests, new signal and strategy-template libraries, more diverse research targeting, and fixes to Live Market data, the liquidation feed, and recurring data-pipeline crash loops.
Leverage support and full multi-venue routing added to strategy backtests
The backtesting engine now applies configurable leverage per strategy, with per-venue caps (up to 20x on futures, capped at 1x on spot and options since those cannot borrow). Strategy routing was also expanded from effectively Binance-only to all seven supported venues, including Binance COIN-M, Bybit, Hyperliquid, Binance Options and Deribit, with guidance for choosing the right venue for options-based strategies.
New indicator and strategy-template libraries for strategy research
Two new reference libraries are now available to the strategy-building agents: a signals library covering 50+ technical, crypto-specific and cross-venue indicators, and a strategy archetypes catalog with end-to-end templates for trend following, mean reversion, grid trading, basis and funding-rate arbitrage, pairs trading, covered calls, microstructure and liquidation-cascade strategies, plus supporting options-pricing helper code.
Strategy research now spreads across more venues, directions and holding periods
Strategy proposals had become heavily concentrated in one venue and direction. Diversity targets now steer proposals toward a wider mix of venues (including newly-enabled Bybit), long/short direction, single- vs multi-symbol scope and holding periods, and those targets are now computed from the live current portfolio instead of a fixed snapshot that could go stale.
Fixed strategies being wrongly abandoned or stuck during coding and data checks
A queueing bug in the pipeline misread normal queue-wait time as a stuck coding agent, causing strategies to be abandoned even though the agent had completed its work correctly. The data-quality gate also no longer blocks a strategy forever once a data gap is confirmed permanently unavailable, and Hyperliquid data refills now finish instead of timing out on large history requests.
Fixed missing data on Live Market futures panels and the liquidation feed
The Live Market page's futures order book, trades, ticker panels and liquidations tape were silently showing no data because of an outdated Binance WebSocket path; spot panels were unaffected. The same endpoint was corrected for the backend liquidation collector, which was also moved to run as its own dedicated process so it stops dropping its connection under load.
Improved reliability of the data-collection and auditing pipeline
Several issues that caused the data-collector and auditor services to crash-loop or silently stall inside their containers were fixed, including lock handling that misbehaved on container restarts and a heartbeat check that could miss a stalled process for days. Historical data auditing coverage was also extended to more data types used in research and backtesting.
Data-integrity pipeline, COIN-M market data fix, and agent reliability improvements
This week's changes add automated data-integrity checks that gate backtests on complete data, fix COIN-M futures data collection, and resolve two bugs causing spurious agent pauses and false backtest failures.
Automated data-integrity checks now gate backtests
A new auditing process continuously scans historical market data for gaps, duplicate rows, and stale writes, recording results for each dataset. Backtests are now blocked from running on any data that fails these checks, and a companion refill process automatically fetches missing data from upstream sources to close detected gaps. This reduces the risk of strategies being validated against incomplete price history.
Fixed COIN-M futures market data collection
COIN-M futures kline data (regular candles, mark price, index price, and premium index) for settled quarterly contracts such as ADAUSD and BCHUSD and active perpetuals such as BTCUSD_PERP and ETHUSD_PERP could not be downloaded because requests were sent to the wrong futures endpoint, causing these gaps to never fill. This data now downloads and backfills correctly.
Fixed spurious agent pauses shortly after platform startup
The coding agent used in strategy research could begin failing every request within minutes of the platform starting, eventually pausing itself after repeated consecutive failures, due to an internal event-loop configuration conflict triggered when other components reinitialized in the background. This conflict is now avoided, keeping the agent available right after startup instead of needing manual intervention.
Backtests no longer fail over unrelated supplementary data
Backtests and optimization runs were checking a strategy's data completeness against every supplementary dataset loaded for an instrument, rather than only the fields the strategy actually reads, which could fail otherwise-valid strategies over data they never use. Completeness checks are now filtered to the supplementary fields a strategy's code actually references before backtesting or optimizing it.
Weekly Update: Internal Repository Housekeeping
This week's only recorded change was an internal source-control fix with no user-facing product impact; no new features, improvements, or fixes are being published.
A quiet week
No product changes were recorded this week.
Paper Trading, Live Market Data, AI Chat Assistant, and Platform Security
This week added paper trading execution, a live market data page, a chat assistant for hypothesis ideation, and HTTP authentication, plus fixes to options data collection and catalog scanning.
Paper trading execution engine
Strategies can now be run in simulated paper trading. A supervisor process manages per-strategy worker processes, and a broker simulation applies exchange-style fees when filling orders, giving a dry-run of strategy performance ahead of any live trading.
Live Market page
A new Live Market page shows real-time candlestick charts, order book depth, live trade prints, strategy signals, and futures metrics for tracked instruments, giving a single view of current market conditions alongside strategy activity.
Chat assistant for hypothesis ideation
A new Chat page lets users converse with an AI assistant to brainstorm and refine trading hypotheses before they are turned into coded strategies, adding an interactive entry point ahead of the automated research pipeline.
HTTP Basic Auth on the web server
The web server now requires HTTP Basic Authentication on every route, including WebSocket connections, closing off unauthenticated access to the dashboard and its live data streams.
Options and Deribit data collection reliability
Fixed the Binance options klines parser to correctly handle both response formats it can return, added the missing expiration parameter required by the Binance options open-interest endpoint, and switched the Deribit collector to a direct connection, resolving 407 errors that were blocking data collection.
Catalog scanning no longer exhausts memory
Instrument existence checks now use a lightweight filesystem lookup instead of loading full catalog metadata, fixing an out-of-memory failure that occurred when scanning catalogs containing tens of thousands of options data files.
Stratmill Weekly Update: New Strategy Dashboard, Chat Assistant, and Pipeline Reliability Fixes
This week adds a strategy monitoring dashboard with an AI chat assistant, expands historical data collection, and fixes several reliability issues including backtest data precision errors and memory leaks in data collection services.
Strategy dashboard with detail view and chat assistant
The dashboard now includes a strategy detail page showing the analysis section and backtest review metrics, with a fallback to the previous iteration's data when needed. A new chat page lets users interact with an assistant to develop strategy ideas, with responses streamed live and submitted ideas rendered as formatted markdown alongside clearer stage-name labels.
More resilient pipeline execution and cross-platform support
The research pipeline now treats certain subprocess initialization failures as transient and retries automatically instead of abandoning the run, and system errors are tracked rather than immediately failing the strategy. Paused agents are now skipped correctly on timeout, and duplicate task creation is prevented, with subprocess health tracked via heartbeats. Process-liveness checks now use a cross-platform library instead of a Windows-specific fallback, improving compatibility across operating systems.
Broader historical data collection and validation
Automatic data collection now pulls the maximum available history from supported exchanges, and data span checks now validate actual calendar-day coverage rather than just bar counts, catching gaps that bar counts alone could miss. Walk-forward optimization results now retain per-window in-sample and out-of-sample Sharpe ratios for review.
Backtests no longer fail on mismatched price/volume precision
Historical bars, trade ticks, quote ticks, order book deltas, and price updates could carry a different decimal precision than the configured trading instrument, causing the backtest engine to reject the data at runtime. Data is now checked and automatically corrected to the instrument's precision before backtests run, preventing avoidable backtest failures.
Resolved memory leaks and duplicate-process issues in data services
Data collection and backtest services were leaking memory over time due to creating new database connections on every call rather than reusing a shared connection, causing steadily growing resource usage. A lock now also prevents duplicate data-collector instances from running simultaneously, and backtest runs explicitly release supplementary data after use.
A quiet week
No product changes were recorded this week.
A quiet week
No product changes were recorded this week.
Automated strategy verification, optimization, and multi-agent pipeline
This week added automated strategy verification and optimization stages, a six-agent research-to-promotion pipeline, new market data collectors, a corrected Sharpe ratio calculation, a pipeline monitoring API, and stability fixes.
Automated multi-layer strategy verification before promotion
New strategies now pass through automated verification before advancing: static code analysis, synthetic market-scenario tests, a sandbox backtest, and QA agent review. Each layer returns specific feedback so strategies that fail a check can be revised rather than moving on to backtesting untested.
Walk-forward and holdout optimization checks for overfitting
Strategies now go through parameter sensitivity analysis to flag settings that only work at exact values, followed by walk-forward and holdout testing on data the strategy has not seen. Strategies whose performance degrades too much out-of-sample or looks brittle to small parameter changes are automatically abandoned rather than promoted.
Full six-agent pipeline from research idea to promotion decision
Trading ideas now move through a chain of specialized agents: a Research Lead and Strategy Developer turn a hypothesis into code, a QA Agent and Strategy Analyst review and evaluate results, and a Risk Officer and Portfolio Manager make the final promotion or capital-allocation call. Agent sessions persist so work can resume after interruptions.
Expanded market data: funding rates, open interest, liquidations
The data pipeline now collects funding rates, open interest, and liquidation data alongside price data, giving the research agents more market context to work from. Data collectors also gained rotating proxy support to help data collection keep running reliably.
Sharpe ratio now calculated from daily returns
Backtest Sharpe ratios were previously computed from per-trade PnL, which could overstate risk-adjusted performance for strategies that make many small trades. The calculation now uses daily returns instead, giving a more accurate performance figure on backtest reports.
Monitoring API for pipeline health and status
A new monitoring API exposes endpoints for system health, pipeline status, active runs, and recorded errors, plus an override endpoint, so the state of in-progress strategy research can be checked and stages can be manually overridden when needed.
Pipeline reliability fixes for stuck stages and agent output parsing
Fixed a bug where a strategy stuck in the coding stage after a restart would never be picked up again; the orchestrator now checks all active stages each cycle. Agent responses wrapped in markdown or surrounded by extra text now parse reliably instead of failing, and database enum values are now written consistently.