Real Time Trading Signals

Size: px

Start display at page:

Download "Real Time Trading Signals"

Cory Jerome Dennis
5 years ago
Views:

1 Real Time Trading Signals Robert Almgren Kx25 May 18,

2 2 Quantitative Brokers business Algorithmic trade execution and cost measurement Agency only: no prop trading or market making Futures and fixed income Equities, FX already well served Global futures exchanges, all products Cash Treasuries (smart order routing) Multi-leg trades: futures vs futures, futures vs cash Metric of success is execution cost relative to benchmark

3 3 Summary of talk Good execution results come from reliable and efficient technology understanding of market structure ability to make short-term price predictions We implement real-time trading analytics in Kdb+ the only external vendor product we use

4 BUY 221 ESM8 BOLT CME S&P 500 futures Passive fills Cumulative exec Market trades Limit orders Cumulative VWAP Microprice Bid-ask Exec = Cost to strike = tick = -$8.51 per lot Market asks Market trades Sweep Strike Benchmark = "Strike" or "Arrival Price" ESM8 WHEAT FEED RICE OUTLOOK 1,000 lots 14:00:13 Our limit orders Market bids Our fills Done at 14:02:22 VWAP Exec Our cumulative average execution price 14:00:00 14:00:20 14:00:40 14:01:00 14:01:20 14:01:40 14:02:00 14:02:20 14:02:40 CDT on Mon 14 May 2018

5 5 SELL 18 UBM8 STROBE Exec = Cost to VWAP = /32 = -$8.65 per lot Strike Sweep Benchmark = "Volume-Weighted Average Price" VWAP Exec VWAP :20 13:40 14:00 14:20 CDT on Mon 14 May :40 15:00:00 13:00 CROP PROGRESS 12:40 Done at 14:51: lots WHEAT FEED RICE OUTLOOK UBM8 Settle :00: Passive fills Cumulative exec Limit orders Cumulative VWAP Microprice Bid-ask CME "Ultra" bond futures 15:00 15:20

6 Multi-leg RM8 Strike Sweep lots 02:23:15 CME 10-year Treasury futures sweep Strike :22:30 02:23:30 02:24:30 02:25:30 02:26:30 02:27:30 02:28:30 (Chicago) SELL 191 RM8 LEGGER Exec = Cost to strike = tick = per lot Done at 02:25:23 VWAP Exec CDT on Fri 11 May 2018 BUY 306 ZNM8 LEGGER 02:23:15 BUY 306 ZN SELL 191 R Exec = Cost to strike = = per lot, total Overall negative slippage 02:22:30 02:23:30 02:24:30 02:25:30 02:26:30 02:27:30 02:28:30 CDT on Fri 11 May 2018 MISS HIT Done at 02:28: Strike Exec Slippage per lot to strike LIFFE Long Gilt futures SELL 191 RM8 LEGGER (London) ZNM8 Sweep Strike ¼ 5,000 lots 02:23:15 Exec = Cost to strike = 0.25 /32 = $7.81 per lot Take leg risk on this leg for passive fills Done at 02:28:12 Exec Positive slippage (Bad) VWAP RM8 Strike Sweep lots 02:23:15 Complete this leg early Exec = Cost to strike = tick = per lot Done at 02:25:23 VWAP Exec Negative slippage (Good) 02:22:30 02:23:30 02:24:30 02:25:30 02:26:30 02:27:30 02:28:30 CDT on Fri 11 May :22:30 02:23:30 02:24:30 02:25:30 02:26:30 02:27:30 02:28:30 CDT on Fri 11 May 2018

7 7 Live and die by transaction costs "slippage" April 2013 through Feb ,512 orders 60.5 lots avg 27,249 orders 58.4 lots avg $/lot 1,724,316 lots 1,591,694 lots $9.5MM in 34 months: 77 bp annual improvement in return

8 Determinants of slippage Passive fills Cumulative exec Market trades Limit orders Cumulative VWAP Cointegration Microprice Bid-ask BUY 573 GEZ8 BOLT Exec = Cost to strike = tick = -$6.25 per lot Limit Sweep Strike BML 500 lots $12.50 Passive fills VWAP Exec buy at bid, sell at ask be patient, unless price will move away GEZ8 100,000 lots 09:43:21 wait 32 minutes until limit orders fill Done at 10:15:49 09:40 09:45 09:50 09:55 10:00 10:05 10:10 10:15 10:20 CDT on Wed 09 May 2018 Overall direction of market BUY $7MM CT10 BOLT hard to predict Passive fills Cumulative exec Market trades Limit orders Cumulative VWAP Microprice Bid-ask Exec = Cost to strike = /32 = -$ per lot Sweep Strike 98-12¾ *** Short term pricing signals price will go up or down? pick when to execute CT10 Market comes down through us $100MM 08:36:04 Done at 08:38:41 Exec VWAP :35:40 08:36:20 08:37:00 08:37:40 08:38:20 08:39:00 EDT on Wed 16 May 2018

9 Trading architecture 9 Parent Client Match engine order book order book order book Child orders and messages Gateways Algorithmic Execution Engine (Proprietary C++) orders and fill reports Gateways Client Client Client order book order book order book Gateways Market data Signal generator (Kdb+) Client The signal generator receives market data, Logs (Kdb+) performs computations to predict prices, and feeds the results to the algorithmic engine to improve trade execution. Exchange Algorithmic broker External

10 10 Data volumes (top of book) 150MM messages / day ~ 2000/sec on average Depth data is order of magnitude larger

11 11 Example signals Prices overshoot and relax Related assets tend to move together Presence of imbalance in the market (large traders)

12 12 Quote imbalance signal "Microprice" = average of bid and ask, weighted by quote sizes Tends to move continuously, as quotes are withdrawn (or trades consume liquidity) before a price change ,000 lots BAXZ8: Dec 2018 Canadian Banker's Acceptance Related signals: trade imbalance (sells - buys) large trades round number trades Indicate presence and desires of other traders in market 10:00 10:05 10:10 10:15 10:20 10:25 10:30 10:35 10:40 10:45 10:50 10:55 11:00 EDT on Mon 09 Apr 2018

13 13 Correlated assets High Dimensional Machine Learning for Interest Rate Regime Change Identification Kyle Xiao Advisor: Robert Almgren Submitted in partial fulfillment of the requirements for the degree of Figure 1: The prices of all five futures studied in this paper are plotted against time. The prices have been centered and standardized. We see that these scaled prices move almost perfectly in sync with one another across the entire day. Bachelor of Science in Engineering Department of Computer Science Princeton University May 2018 (a) Plot of Ten Year Price vs Five Year Price with PC1 (b) Plot of Ten Year Price vs Bond Price with PC1 Figure 2: We see that in both plots the prices primary deviate along PC1, the line drawn in grey.

14 14 Cointegration example BUY 9 IZ0 BOLT Equilibrium price Aggressive fills Passive fills Cumulative exec Market trades Limit orders Cumulative VWAP Cointegration Microprice Bid-ask Exec = Cost to strike = 0.72 tick = 9.03 per lot rises above offer: we aggress (wrong in this case) Exec Sweep VWAP Strike IZ0 10,000 lots 07:12:03 IZ0: Dec 2020 Euribor (LIFFE) Done at 07:49:20 07:05 07:10 07:15 07:20 07:25 07:30 07:35 07:40 07:45 07:50 07:55 BST on Fri 11 May 2018

15 15 Cointegration projection computation in Q Cointegration is a singular value decomposition We need only first vector, use power method while[ (err > tol)&(not vold ~ neg v)&(it<=500i); vold:v; tempv:covma mmu v; v:tempv % sqrt sum tempv*tempv; err:sum (vold-v)*vold-v; it:it+1; ];

16 16 Sweep (reversion) signal BUY 23 GCQ4 BOLT Aggressive fills Passive fills Intended passive Cumulative exec Market trades Limit orders Cumulative VWAP Microprice Bid-ask Exec = Cost to strike = tick = -$48.04 per lot Sweep Strike To make this work: condition on several other variables describing VWAP Exec market state GCQ4 100 lots 09:52:38 Sharp motion up followed by forecast Done at 09:58:49 of reversion down to specific level until specific time 10:07:38 09:42 09:46 09:50 09:54 09:58 10:02 10:06 CDT on Mon 28 Jul 2014

17 Intraday bubbles Example Buy Signal The Detection of Intra-Day Bubbles Test is a generalized version of Augmented-Dickey Fuller test of unit root The prototypical model takes the following form:

17 17 Intraday bubbles Example Buy Signal The Detection of Intra-Day Bubbles Test is a generalized version of Augmented-Dickey Fuller test of unit root The prototypical model takes the following form: Momentum signal The market was trending up Our model correctly identified this and produced a signal about 2 minutes after the rally started (around 2:39 am) The signal expired after the price flattened out (around 2:44 am). When the price is believed to be in an explosive state. quantitativebrokers.com To make this work: condition on several other variables describing market state Shankar Narayanan, quantitativebrokers.com Quantitative Brokers

18 18 Smart Order Routing Logistic regression Random forest regression Multilayer perceptron Gradient boosting regression

19 updatemavgprice.z.ts updatesvd fetchu 19 Q signal architecture updatequotesize calc_avgquotesize get_avgquotesize checkbq4newswp subcheckbq4newswp checkbq4swp checkbq4weakswp subcheckbq4weakswp upd updatemidprc getrtvol Incoming market data updatemidhist updatemidhistwdrop updatemidhistwodrop calc_tevavgtradeprc update from C plugin updatebvlm get_tevavgtradeprc weakswptest newswpacchelper newswptest publish2tp Outgoing data to algos updatesvlm weakswpgen newswpacc subswpinv via C plugin processbestquote weak sweep newswpinv processtrade updateprojections cointegration checkspectime swpgen swpinv swpacc swptest addbswp snapshot trdfilt sweep addsswp updatemindt siggen rginv updaterg updatevlmprc range rgfinv vlmtrd updatesvlmsig updatebvlmsig

20 20 Conclusions High frequency trading is computationally demanding Short-term price prediction is key to performance Kdb+ for real-time data processing

TCA what s it for? Darren Toulson, head of research, LiquidMetrix. TCA Across Asset Classes

TCA what s it for? Darren Toulson, head of research, LiquidMetrix We re often asked: beyond a regulatory duty, what s the purpose of TCA? Done correctly, TCA can tell you many things about your current