Loading
Edge Feed
Preparing ranked opportunities for investigation.
Loading opportunity
Fetching evidence-linked feed data.
Loading opportunity
Fetching evidence-linked feed data.
Loading opportunity
Fetching evidence-linked feed data.
Preparing ranked opportunities for investigation.
Fetching evidence-linked feed data.
Fetching evidence-linked feed data.
Fetching evidence-linked feed data.
Months of accumulated learning, made visible without being confused with trade decisions.
Research history — not trade recommendations
This is the durable record of every alpha idea SharpTrace has tested, from the same registry every alpha sprint updates by hand. Nothing on this page is a BUY/WATCH/PASS decision.
Status Summary
0
ACTIONABLE4
SHADOW11
OBSERVING19
CLOSEDMultiple-Testing Context
33 materially distinct alpha hypotheses have been tested to date (excludes infrastructure/coverage work). Treat any single marginal pass with appropriate skepticism - at this trial count, some fraction of apparently positive results are expected by chance alone even if no real edge exists.
2
TRADER COPY
5
INFORMATIONAL LATENCY
2
FORECASTING
2
STRUCTURAL RELATIVE VALUE
10
MICROSTRUCTURE
10
CROSS VENUE
2
LIQUIDITY PROVISION
1
INFRASTRUCTURE
A category with a high count has been heavily tested (over-tested risk); a category absent or near-zero here remains underexplored.
Research Throughput
Decision quality, not hypothesis volume - “hypotheses/month” is deliberately not tracked here.
Insufficient historical metadata
Median days: predeclaration → decision
Insufficient historical metadata
Killed before prospective implementation
Insufficient historical metadata
Historical → prospective survival rate
0
Strategies graduated
The registry records one version-label per row, not a predeclaration/decision date pair, and pre-v37 rows do not reliably state whether prospective deployment happened before rejection - both are reported as insufficient rather than reconstructed from guesses. Strategies graduated is a real count from shadow_strategy_registrations.
Active — being watched or validated
Sports Catalyst V2
Progress frozen at 98 events (85 direction-assigned); economic verdict **PARTIALLY** (see v36.3 decision report)
Last updated v36.3 · `V36_3_DECISION_REPORT.md`
Wunderground weather catalyst
Progress 6 real point-in-time cases, station-correspondence risk confirmed real (NYC KLGA≠KNYC), decision **INSUFFICIENT_EVIDENCE** pending more cases
Last updated v36.3 · `V35_1_WEATHER_HISTORICAL_FALSIFICATION.md`, `V36_3_DECISION_REPORT.md`
Kalshi↔Polymarket weather historical price discovery
Closed — rejected, parked, or completed
Kept separate from the active list above so a closed idea is never mistaken for a live signal.
Global trader skill
Last updated pre-v34.0 · earlier session
Crypto trader skill
Last updated pre-v34.0 · earlier session
Broad Steam detector
Outcome storage disposition closed v41.3A: `steam-microstructure-collector`/`steam-detection-tracker` containers were already stopped (Exited 4 days), `run_steam_detection_tracker.py` given an explicit `RuntimeError` guard so it can't be silently restarted against dropped tables, legacy `orderbook_snapshots`/`orderbook_levels` (411MB, last write 2026-08-17, superseded by `kalshi_orderbook_snapshots`) dropped. This is the case study behind the v41.4 experiment-lifecycle policy — the dead infrastructure sat unnoticed for days behind two still-referenced tables until a storage audit found it.
Last updated pre-v34.0; storage closed v41.3A · earlier session; `DATA_RETENTION_AND_STORAGE_BUDGETS.md` §3, §6
Progress evidentiary weight downgraded — the v38.1 raw signal (406 pairs, 12,309 aligned obs, 65.5% positive edge) was forensically audited in v39.0 Track A: 14.5% of observations (30.9% of total gap magnitude) are `OUTCOME_EFFECTIVELY_KNOWN` contamination — Kalshi's price already reflects the settled outcome while the joined Polymarket quote predates that information (same failure class as the sports 21:2→3:1 collapse). Signal survives filtering (≥5pt rate 51.6%→45.5%, positive-edge rate 67.3%→62.1%) but the raw numbers must not be cited as-is. A real settlement-source regime change (Kalshi NWS→The Weather Company, Aug 14) also spans most of the historical sample. v37.0-style prospective shadow now deployed (see below) as the authoritative test
Last updated v38.1, v39.0 · `V38_1_NON_SPORTS_COMPARATIVE_DATASET.md`, `V39_0_TRACK_A_FORENSIC_AUDIT.md`
Kalshi-leads/lags-Polymarket weather prospective shadow (v39.0)
Progress `CLOSED / REJECTED` — frozen analysis run at 29 fully-observed episodes (V39_1). Reason: the live-uncertainty window this hypothesis needed to observe never survived past signal-open (0/29 episodes) - a data-availability finding, not a "prices don't converge" result. `weather-lead-shadow` service left running (collector preserved), no retuning
Last updated v39.0, v43.2, v43.4, v39.1 · `V39_0_TRACK_A_FORENSIC_AUDIT.md`, `V43_2_STRATEGY_GRADUATION_BOARD.md`, `V39_1_WEATHER_FROZEN_ANALYSIS_VERDICT.md`
View live SHADOW signals →Polymarket↔Kalshi arbitrage (live)
Progress `CLOSED / REJECTED — NO EXECUTABLE EDGE OBSERVED AT SUFFICIENT FREQUENCY`. Median edge -$0.017 (331 events) already showed no reliable pattern. v43.4 forensic audit of all 44/331 (13%) events with `dir1_net_worst_case` >$0.05: 22 were already self-flagged by the scanner's own v36.0 sanity check as side-mapping artifacts (net outside the plausible [-0.50,0.10] band). The remaining 22 (all `mapping_confidence=EXACT`, so team/game matching itself was fine) decompose to the SAME root cause under manual audit: the "NO-home Kalshi" leg was sourced from the AWAY team's Kalshi contract instead of the home team's in the majority of cases, across every affected pair (MLB:ATH@KC, TOR@TB, DET@KC, MIN@SD, CIN@AZ) - a side-assignment bug, not a real cross-venue mispricing. All 44 high-edge observations are attributable to known mapper defects; none survive audit. Closed per instruction rather than extending observation indefinitely. Data preserved; scanner left running (288KB, negligible) for whatever future mapper-fix work might reuse it, but no further active development on this hypothesis
Last updated v36.1, v43.3, v43.4 · `V36_1_MULTIVENUE_FALSIFICATION.md`, `V43_3_UNBLOCK_STRATEGY_PIPELINE.md`, `V43_4_GENERAL_ARB_FORENSIC_AUDIT.md`
Kalshi→Polymarket historical price discovery
Progress evidentiary weight downgraded — the original 21:2 result was re-run from scratch post-mapper-fix under strict rules (EXACT-only, ≤15min freshness as primary, true pregame): collapses to 3:1 (n=6, 5 distinct games). Directionally consistent with Kalshi leading, but n is too small to be independent evidence — v37.0's live prospective test is now the authoritative test, not this reconstruction
Last updated v36.3, v37.1, v37.2 · `V36_3_DECISION_REPORT.md`, `V37_1_HISTORICAL_LARGE_GAP_AUDIT.md`, `V37_2_MAPPER_FIX_AND_REAUDIT.md`
Kalshi-leads-Polymarket prospective shadow (v37)
Progress unconditionally with this reasoning (see `V37_REJECTED_REASON` in `strategy_validation.py`), regardless of any signal count observed after 2026-09-02. Collector left running (data/infrastructure preserved, consistent with how Weather-v39 and Sportsbook→Polymarket were closed) but no further active development: no retuning, no extended deadline, no reopening on later counts
Last updated v37.0, v37.1, v43.3, v43.5, v47.1, v48.4 · `V37_0_KALSHI_LEAD_SHADOW_CHECKPOINT.md`, `V37_1_HISTORICAL_LARGE_GAP_AUDIT.md`, `V43_3_UNBLOCK_STRATEGY_PIPELINE.md`, `V43_5_V37_SIDE_FIX.md`, `V48_4_ROADMAP_RESET.md`
Sportsbook→Polymarket price discovery
Progress `CLOSED / REJECTED` — frozen N=40 primary cohort evaluated exactly per `V43_3_SPORTSBOOK_SUCCESS_KILL_SPEC.md`: mean convergence +0.0256 (pass) and 90% CI [0.0120, 0.0412] excludes zero (pass), but win rate 32.5% decisively fails the ≥60% success bar (and the 50% failure floor). Real average price movement toward sportsbook consensus exists but is not a reliable per-case edge. Collector left running (data/infrastructure preserved: 26 pre-boundary + 111 post-boundary cases as of 2026-09-07); no further threshold tuning, subgroup creation, or "Sportsbook v2" — closed per instruction rather than retuned
Last updated v36.1, v36.3, v43.3, v43.4, v43.7, v46.1 · `V36_1_MULTIVENUE_FALSIFICATION.md`, `V36_3_DECISION_REPORT.md`, `V43_3_SPORTSBOOK_SUCCESS_KILL_SPEC.md`, `V43_7_SPORTSBOOK_COLLECTOR_FIX.md`, `V46_1_SPORTSBOOK_FROZEN_VERDICT.md`
View live SHADOW signals →Kalshi commodity market-making (fill/adverse-selection)
Progress `CLOSED / REJECTED — PASSIVE SPREAD CAPTURE DOES NOT SURVIVE MULTI-SESSION VALIDATION`. v43.3 confirmed on 48h/multi-session data (n=2,231-2,505 fills/series, ~20x the original ~2h sample) - discovered collection never stopped (general Kalshi ingestion has run continuously since 2026-08-20; the "2h" was the original analysis window, not the collection lifespan). Positive rate held at 36-49% across WTI/GOLD/SILVER at both +60s and +5m, still negative-to-breakeven net-of-fee at every series/horizon - the larger sample confirmed rather than reversed the original finding. v43.4: closed per instruction - data/infrastructure preserved (the general Kalshi ingestion tables this drew from keep running for other purposes), but no further active strategy development: no catalyst guard, no maker bot, no quoting optimization
Last updated v38.1, v40.0, v43.3, v43.4 · `V38_1_NON_SPORTS_COMPARATIVE_DATASET.md`, `V40_0_KALSHI_MM_ECONOMICS.md`, `V43_3_UNBLOCK_STRATEGY_PIPELINE.md`
KXBTCD prospective observation (capacity/monotonicity/cold-start)
Progress new lightweight collector `kxbtcd-prospective` deployed v41.1, captures full ladder (top-of-book size, bid/ask, result) at open/+5m/+15m/+30m/+60m/resolution checkpoints per hourly event. Serves three purposes from one capture: capacity-size proxy (Track E), structural-monotonicity forward-repricing (Track H, separate from the rejected calibration strategy above), and cold-start lifecycle (Track I, 5 metrics predeclared before any data existed). No results yet — just deployed
Last updated v41.1 · `V41_1_KXBTCD_ADVERSARIAL_VALIDATION.md`
Kalshi `KXMLBGAME` mispricing vs. no-vig sportsbook consensus
Progress predeclared pregame mispricing test (gap ≥0.05 between Kalshi's home-side price and `sportsbook_reference_cases`' own consensus, run once, no re-tuning after seeing the result). 0 of 20 parseable games survived: 13 had no Kalshi trade before the game's own start time at all, 6 more had their matched sportsbook snapshot taken mid-game rather than pregame, 1 thin-books, 1 no-match. The market itself does not trade meaningfully pregame - a structural data-availability finding, not a methodology fix. Cross-sectional Kalshi longshot-calibration generalization (the other v46.3 candidate) was separately deprioritized, not tested — its historical population (KXINXU/KXBTCD) has aged out of `kalshi_markets`' retention window
Last updated v46.3, v46.4 · `V46_3_NEXT_ALPHA_AND_THESIS_AUDIT.md`, `V46_4_TRADING_SIGNAL_ENABLEMENT.md`
ESPN pregame win probability vs. Polymarket executable price (SPORTS fair value)
Progress the field the codebase originally targeted (`predictor`) is absent from 100% of 244 completed-game summaries checked (live/upcoming-game-only, never historically queryable - explains its prior zero callers). The real historical field (`winprobability[0]`, confirmed genuinely pregame) shows real but weak calibration against actual outcomes (Brier 0.2439 vs. 0.25 uninformative baseline, n=244, MLB only - NFL/NBA/NHL off-season this window) - not a clean pass. Predeclared independent-value test against Polymarket's own executable price (never sportsbook consensus): n=82, win rate 46.3% (below the 50% floor), 90% CI [-0.0706, 0.1176] includes zero. ESPN's pregame number carries no exploitable edge against an already-efficient Polymarket price. Separately identified (not fixed): the `FORECAST_MODEL_AVAILABLE` discovery channel's unordered `LIMIT 500` query on a 65,520-row eligible pool makes newly-crypto-extended markets (~0.3% of that pool) nearly invisible to discovery - a likely-unintentional gate, diagnosed only
Last updated v46.6 · `V46_6_ESPN_FAIR_VALUE_FALSIFICATION.md`
Kalshi 15-minute Gold/Silver entry-price calibration bias (fade YES near 50/50)
Progress emerged from a coarse 3-family exploratory scan (Gold/Silver/WTI 15-minute commodity threshold markets), not a predeclared idea; frozen and tested once before acting on the exploratory count. Primary: Gold N=58 win rate 50.0% mean net edge -0.0116, Silver N=55 win rate 50.9% mean net edge +0.0022, both 90% CIs include zero. Predeclared negative control (WTI, N=50) outperformed both hypothesis assets (56.0% win rate, +0.0344 mean edge) - the opposite of what the hypothesis predicts, confirming the original exploratory bucket count was noise, not a structural regularity
Last updated v46.9 · `V46_9_TRADEABLE_POPULATION_SEARCH.md`
Kalshi `KXMLBGAME` in-game price lag vs. ESPN win-probability jumps
Progress (at the feasibility gate, no P&L calculated) — predeclared before any query: does Kalshi's live price lag a material (≥0.15) ESPN win-probability jump, using ESPN's own play-by-play as an independent, non-sportsbook game-state source. Feasibility check found only 16 matchable game-state observations (a real Kalshi trade landing 30-600s after the event), clustered into just 5 independent games, despite 35,048 real in-game Kalshi trades existing across those same tickers (liquidity was never the bottleneck - timing coincidence between a large jump and a nearby trade is simply rare at this sample size). Too weak and clustered to support any inference; not rescued with new infrastructure, per instruction
Last updated v47.0 · `V47_0_LAST_MILE_SIGNAL_SEARCH.md`
ESPN pregame win probability vs. Kalshi's own pregame executable price (`KXMLBGAME`, no sportsbook consensus anywhere)
Progress distinct from both prior ESPN tests (row 63 tested vs. Polymarket's price; row 62 tested Kalshi vs. sportsbook consensus): this is ESPN's pregame number used directly as fair value against Kalshi's own liquid, settlement-clean price. Predeclared, run once: of 881 ESPN games with a real pregame win-probability reading this season, only 4 had both a matching Kalshi ticker and a real trade inside a tight [-30,+5min] pregame window (KXMLBGAME's real liquidity is heavily in-game, not pregame - consistent with row 65's 35,048 in-game trades). N=4, win rate 75%, mean net edge +0.2375, 90% CI [-0.025, 0.500] includes zero. Killed by the letter of the predeclared bar, but honestly a sample-size failure, not a demonstrated absence of edge - the raw numbers on this thin sample look favorable
Last updated v48.1 · `V48_1_MLB_FALSIFICATION_AND_COMMODITY_FEASIBILITY.md`
BTC threshold model
Outcome settlement incompatible
Last updated pre-v34.0 · earlier session
Short-dated equity (v32.1 lognormal)
Last updated pre-v34.0 · earlier session
Threshold-ladder relative-value arbitrage
Outcome 0/70 executable violations survived
Last updated v34.1 · `V34_1_PARALLEL_ALPHA_TESTS_CHECKPOINT.md`
Pregame MLB favorite-longshot bias
Outcome insufficient tail population (0 obs below 20%/above 80% at kickoff, n=396)
Last updated v36.3 · `V36_2_CASH_IN_THE_KALSHI_DATA.md`, `V36_3_DECISION_REPORT.md`
Sports Catalyst V1
Outcome 25-event frozen cohort, superseded by V2
Last updated v35.0 · `V35_0_SPORTS_V1_FREEZE.md`
Kalshi↔Polymarket weather station equivalence
Outcome deterministic station-identity mapper built and empirically verified against each venue's own stated station code (never inferred from city name). 12 cities checked: 8/12 (66.7%) empirically equivalent, 4/12 real mismatches (NYC KLGA≠KNYC now confirmed with exact mechanism, plus 3 newly-found: Dallas KDAL≠KDFW, Denver KBKF≠KDEN, Chicago KORD≠KMDW)
Last updated v38.1 · `V38_1_NON_SPORTS_COMPARATIVE_DATASET.md`
Kalshi↔Polymarket Fed rate mapping
Outcome 0 exact pairs across all 3 checked non-emergency Polymarket proposition shapes (cumulative-by-meeting: DECISION_DATE mismatch; per-meeting bps-delta: OUTCOME_SEMANTICS mismatch — the literal "Fed cuts vs Fed cuts 25bps" case; end-of-year snapshot: DECISION_DATE mismatch), vs Kalshi's absolute-rate-level-per-meeting `KXFED` series (22 settled markets, 8 meetings). Structural incompatibility, not a mapper bug — do not re-test without a genuinely new Polymarket Fed market shape appearing
Last updated v38.1 · `V38_1_NON_SPORTS_COMPARATIVE_DATASET.md`
Historical executable Polymarket↔Kalshi arbitrage
Outcome the reported 25/23 positive-edge episodes were contaminated (65% wrong-contract-type matches via a now-fixed mapper bug, remainder stale-timestamp artifacts); post-audit survivor count is 1, with net edge = $0.00 exactly. Do not cite as supporting evidence
Last updated v37.1, v37.2 · `V37_1_HISTORICAL_LARGE_GAP_AUDIT.md`, `V37_2_MAPPER_FIX_AND_REAUDIT.md`
Kalshi maker rewards
Outcome real programs exist but 100% concentrated in 3 fifteen-minute commodity series, excludes existing market makers
Last updated v36.1 · `V36_1_MULTIVENUE_FALSIFICATION.md`
KXINXU longshot-overpricing calibration bias
Outcome (as a naive fixed-rule strategy) — v40.1's promising in-sample result (4,290 contracts/10 days, positive rate 78-81%, mean P&L +$0.07-0.10) did NOT survive v40.2's true out-of-sample adversarial validation (26,376 contracts/252 events/36 independent trading days, zero overlap with training window). Out-of-sample: win-rate direction replicates (65-69% positive, CI clearly >50% even clustering by day) but **mean/total P&L is negative at both T-60m and T-30m**, driven by a large loss tail (mean loss -$0.67 vs mean win +$0.27 — needs >71% win rate to break even, came in at 65-69%). Worst 5% of trades alone exceed the entire realized loss. Real instability found too: H1000/H1200 were the strongest hours in-sample, became net losers out-of-sample. Also corrected v40.1's wrong "opaque Kalshi-modified source" characterization with real primary documentation (original 2022 CFTC filing): source was originally Quandl Data, expiration time 7:00pm ET; current live source still not independently named. No prospective shadow recommended. The underlying calibration/win-rate pattern may still be real (it replicated directionally) but requires a fundamentally different economic treatment (position sizing for the loss tail, or a different implementation) before any deployment consideration — not a naive fixed-size fade rule
Last updated v40.0, v40.1, v40.2 · `V40_0_KALSHI_MM_ECONOMICS.md`, `V40_1_KXINXU_CALIBRATION_BIAS.md`, `V40_2_KXINXU_ADVERSARIAL_VALIDATION.md`
Longshot-overpricing generalization beyond KXINXU (KXBTCD)
Outcome (as a naive fixed-rule strategy) — v41.0's 20-event exploratory read looked directionally similar to KXINXU's own training result; v41.1's true OOS validation (14,135 contracts, 236 independent hourly events, 10 independent days, Aug 1-10, zero overlap with the 20-event sample) found mean P&L statistically indistinguishable from zero (90% cluster-bootstrap CI includes zero both by event and by day: T-60m [-$0.049,+$0.052], T-30m [-$0.053,+$0.054]). Breakeven hit rate (70.0-70.3%) matches the actual win rate (70.3-70.4%) almost exactly. Same tail-loss mechanism as KXINXU confirmed: nearly identical mean-win/mean-loss magnitudes (~$0.28 win vs ~$0.66 loss on both families) despite BTC and the S&P being structurally unrelated underlyings — flagged as a real research observation (a structural feature of "fade a moderate Kalshi longshot net of Kalshi's fee" as a trade shape, not a per-family fluke) worth remembering, not chasing further right now. No prospective shadow recommended
Last updated v41.0, v41.1 · `V41_0_KALSHI_OPPORTUNITY_DISCOVERY.md`, `V41_1_KXBTCD_ADVERSARIAL_VALIDATION.md`
KXBTCD / crypto-hourly family (BTC/XRP/SOL/DOGE/BNB threshold + range ladders)
Outcome inventory only. Discovered in v41.0's market-structure census: materially more liquid than every other family studied to date (KXBTCD alone: $2.4M/24h volume vs KXINXU's $33K), CF Benchmarks BRTI settlement (documented, verifiable methodology), 24/24h recurring cadence, 400+ settled markets. No strategy work started — flagged as the strongest candidate for future study given its liquidity and settlement transparency exceed every other family checked
Last updated v41.0 · `V41_0_KALSHI_OPPORTUNITY_DISCOVERY.md`
Billboard 200 cross-venue pair (Kalshi KXTOPALBUM vs Polymarket weekly #1-album markets)
Outcome real, structurally-matching candidate found in v41.0's cross-venue inventory expansion (same weekly #1-ranking proposition on both venues). No mapper built, no pairs verified yet — settlement-source compatibility not independently checked. US CPI and ADP employment were also checked as candidates and found NOT viable (Polymarket's CPI markets are all China CPI — a country mismatch; no genuine ADP match found) — real negative results, not gaps
Last updated v41.0 · `V41_0_KALSHI_OPPORTUNITY_DISCOVERY.md`
Kalshi non-sports/weather calibration population (KXNATGASD)
Outcome inventory only, no strategy started. `KXNATGASD` (natural gas daily threshold, highest raw volume checked) remains a candidate for the same kind of study now underway on KXINXU. `KXDJI`/`KXPLATINUMH` checked and found dead (zero volume) despite same family shape — liquidity is series-specific, not category-wide
Last updated v40.0 · `V40_0_KALSHI_MM_ECONOMICS.md`
Cross-sectional generalization of Kalshi longshot-calibration pattern (KXINXU/KXBTCD)
Outcome future research item only, per explicit instruction: a calibration *study* (does the same win-rate/loss-tail pattern found independently in both KXINXU and KXBTCD replicate across other unrelated Kalshi families?), not a trading rule and not a rescue of either rejected strategy. No work started
Last updated v42.1 · `V42_1_RESEARCH_GOVERNANCE.md`
Kalshi broad catalog coverage
Outcome full active catalog (~11,200 events, ~3,150 material series) on a bounded rotating cadence; breadth judged sufficient, no further broadening planned absent an operational problem
Last updated v38.0 · `V38_0_KALSHI_COMPARATIVE_INTELLIGENCE.md`