Roadmap

• [x] multi-market support within strategies PR#30, PR#53, PR#54, PR#64
• [x] better position sizing PR#8
• [x] fee modeling PR#4, PR#42
• [ ] fuller slippage modeling for maker realism still needs L3 data PR#6, PR#9, PR#50
• [x] Polymarket L2 order-book backtests PR#10, PR#45, PR#57
• [x] PMXT raw archive workflow: local raw mirrors, direct archive fallback, filtered replay cache, and incremental downloader reruns PR#17, PR#22, PR#40, PR#47, PR#56, PR#61, PR#64
• [x] Rust-native staged data loading for PMXT and Telonex with unified cache/local/archive/API progress output PR#132
• [ ] Kalshi L2 order-book backtests need L2 historical book data we do not have yet. The next exchange expansion targets are Limitless.exchange and Opinion.trade after the Polymarket loading path stays stable.
• [x] richer charting and honest multi-run HTML/report outputs PR#5, PR#52, PR#68, PR#74, PR#80, PR#83
• [x] manifest-based runner architecture and repo-level optimizer surface PR#67
• [x] repo-level runner/report contracts, docs validation, and launcher/docs hardening PR#64, PR#65, PR#68, PR#69, PR#71, PR#76, PR#77, PR#78, PR#80, PR#81
• [x] archived Telonex private research source, BTC snapshot model profiles, and BTC snapshot live sandbox examples published for study PR#147

Known Issues

• The archived backtests/private/ and strategies/private/ source is published for study, not as a current trading edge. These runners capture old Telonex BTC 5m, general-market, and resolved-sports research surfaces. They can require Telonex API access, local Telonex cache coverage, and substantial replay memory. Market structure, public feed behavior, Polymarket constraints, and model relevance have drifted since the recorded research runs.
• The archived BTC snapshot model bundle is the only private strategy family wired into live sandbox runners. Re-training the included JSON profiles requires local Telonex Polymarket book snapshots and Telonex Binance spot parquet caches that are not shipped with the repository; cross-asset profiles need matching ETH, SOL, or XRP spot caches. The archived cache-hydration helpers can populate those paths for users with Telonex API access, but exact reproduction still depends on historical Telonex file availability and matching the original environment settings. The companion walk-forward and runner-validation scripts also require compatible dataset/model artifacts produced by the research workflow.
• Kalshi is not currently exposed as a public runnable backtest path. The repo still contains Kalshi instrument, trade/candlestick loader, fee-model, and research helper components, but the built-in replay adapter registry only exposes Polymarket PMXT and Telonex book replay adapters. Users should treat Kalshi as experimental adapter plumbing until real Kalshi L2 historical book data, a Kalshi replay adapter, and a public runner are added.

Recently Fixed

• [x] v4 staged replay loading now prepares Polymarket metadata first, loads all book data second, and loads execution trade ticks last. PMXT filtered-cache misses are grouped by raw archive hour, Telonex API/file work has separate worker caps, and warm 100-market Telonex loads avoid redundant full-week book sorting/count scans PR#132.
• [x] PR#119 upgrades public Polymarket runners to L2-native BookReplay semantics with OrderBookDeltas plus real TradeTick execution evidence, keeps Nautilus on BookType.L2_MBP with trade_execution=True, removes standalone quote/trade replay framing, speeds Telonex API-day cache reads with .fast.parquet sidecars, prunes local Telonex mirror scans through the DuckDB manifest, bounds the Telonex downloader writer queue, periodically closes open Telonex part writers, and makes PMXT raw downloads incremental by skipping existing local files PR#119
• [x] PR#119 now also materializes Telonex OrderBookDeltas under book-deltas-v1, prints richer terminal statistics from per-market result payloads plus portfolio-level Nautilus BacktestResult stats, and keeps public Python runner inputs inline inside run() instead of module-level constants PR#119
• [x] prediction-market backtests now use settlement-aware result assembly, finite synthetic taker depth, lower default touched-limit fill probability, fill-time Kalshi fee waivers, zero-fee Polymarket maker modeling, safer trade-tick liquidity caps, and corrected Telonex outcome selection; docs, UML, tests, and public runner configs were updated alongside the execution honesty fixes PR#118
• [x] Telonex source loading now falls through to the next source when one fails (so a missing local: mirror cleanly hands off to api:) instead of aborting the whole replay PR#105
• [x] v3 adds a Telonex joint-portfolio runner, local Telonex daily-Parquet downloader, Hive-partitioned Parquet output with a DuckDB resume manifest, and daily-file timing output for Telonex local: / api: sources PR#104
• [x] v3 removes the active PMXT relay path, relay badges, and relay service package. PMXT runners now use local raw files plus direct r2v2.pmxt.dev / r2.pmxt.dev archive fallback, and Telonex is available as a separate Polymarket vendor PR#103
• [x] relay schema-bootstrap now commits its UPDATE so a second writer can take the WAL lock instead of deadlocking on first start PR#102
• [x] PMXT relay latest-hour badge now prints the mirrored filename (polymarket_orderbook_YYYY-MM-DDTHH) instead of only the naked hour, the mirror worker validates the local file size against upstream before reusing an existing raw file (so stale placeholder downloads can no longer survive as ready), count_raw_dump_files excludes undersized parquet files from the public coverage denominator, and startup adoption purges orphan raw files under the nonempty byte threshold that aren't tracked as ready in the index DB PR#101
• [x] PMXT relay archive-coverage redesign exposes per-source priority metrics, mirrored-vs-archive accounting, and clearer empty-hour handling PR#100, PR#95, PR#94, PR#93, PR#90, PR#89
• [x] PMXT runners support full per-entry source ordering and split raw-archive sources, so MarketDataConfig.sources can interleave local: and archive: entries in any order the runner needs PR#98, PR#92
• [x] public runner validation harness covers every flat runner under backtests/ to keep direct script paths and the menu deterministic PR#91
• [x] ruff-driven cleanups across relay, adapters, and plotting; docs now include an embedded-HTML example, acknowledgements, and a fixed index anchor URL PR#88, PR#87
• [x] multi-market runners now default to EMIT_HTML=False and the artifact pipeline downsamples price points to 5 000 before building dense equity curves, cutting wall time from ~320s to ~26s on an 8-market basket PR#84
• [x] HTML chart files are now downsampled to ~5 000 points before Bokeh serialization, reducing a 446 K-bar chart from 31 MB to under 1 MB; redundant ColumnDataSource columns and intermediate DataFrames were also deduplicated, and new regression tests enforce that 100 K-bar backtests produce HTML under 5 MB PR#83
• [x] aggregate summary report builders now skip serializing unused per-market price series, fill events, and overlay curves when the selected summary panels do not render them PR#83
• [x] docs deploy workflow now triggers on the active v2 branch instead of the removed main branch, and the GitHub Pages environment allows v2 deploys PR#81
• [x] plotting docs rewritten around a clearer detail-vs-summary mental model, stale blob/main GitHub links fixed across all docs, and a regression test guards against stale branch links returning PR#80
• [x] backtest runner examples refreshed to match current runner contracts PR#78
• [x] legacy Kalshi trade-tick runners pinned end_time to a known-good close window so direct script paths and the repo pytest gate stay deterministic, the docs/examples now point at current runnable entrypoints, and shared startup reporting no longer understates factory-backed runs PR#76, PR#77
• [x] public runners now use typed replay specs plus one experiment builder, adapter-owned replay loading, and a repo-layer optimizer surface instead of the older shared SIMS / BACKTEST contract PR#67
• [x] plotting now scales as one detailed HTML per loaded sim plus one aggregate summary HTML per basket, the repo no longer relies on concatenated mega-pages, and the prediction-market runner internals are split into clearer execution, artifact, reporting, and data-source seams Issue #73, PR#74
• [x] direct script HTML outputs now resolve from the repo root, fixed-basket multi-market runners emit aggregate reports again, and the repo runner/report surface stays explicit about per-sim detail charts versus aggregate multi-market reports PR#68
• [x] setup/backtest/fetch-source docs now match the unified main.py launcher and current PMXT terminal/reporting output, and the orphaned _trade_tick_ui.py helper is gone PR#69
• [x] root README scope and agent guidance now keep detailed operational docs out of the README body and in docs/ instead PR#70, PR#71
• [x] PMXT L2 replay now orders book updates deterministically so longer windows do not lose book state PR#26
• [x] relay misses fall back client-side to r2.pmxt.dev, trusted proxy clients keep distinct rate-limit buckets, and stale buckets are pruned instead of accumulating forever PR#22, PR#25, PR#42
• [x] relay observability and survivability improved with progress badges, ClickHouse ingest, retry handling around transient lock contention, mirror pruning, and incremental raw-hour adoption PR#34, PR#35, PR#36, PR#40, PR#56, PR#64
• [x] PMXT public workflows are now raw-first: local raw mirrors, archive fallback, mirror-only relay behavior, and downloader output all line up across runners and docs PR#45, PR#47, PR#57, PR#60, PR#64
• [x] public runners now model queue position and static latency where this repo uses them, reducing dependence on zero-latency assumptions PR#50, PR#64
• [x] replay/report outputs now distinguish requested windows from loaded windows and keep honesty-focused defaults visible in normal runs PR#52, PR#56, PR#63, PR#64
• [x] the interactive menu again shows full runner contents, direct runner imports work in both script and package modes, and the root _script_helpers.py shim is gone PR#53, PR#62, PR#64
• [x] PMXT timing output, source labels, and raw-hour progress reporting are clearer and better aligned with the actual runner behavior PR#55, PR#59, PR#60
• [x] repo CI and docs validation now match the documented local gate, and PR docs builds validate without trying to deploy Pages PR#58, PR#64